NBER WORKING PAPER SERIES

MANAGEMENT AS A TECHNOLOGY?

Nicholas BloomRaffaella Sadun

John Van Reenen

Working Paper 22327http://www.nber.org/papers/w22327

NATIONAL BUREAU OF ECONOMIC RESEARCH1050 Massachusetts Avenue

Cambridge, MA 02138June 2016

We would like to thank Orazio Attanasio, Marianne Bertrand, Robert Gibbons, John Haltiwanger, Rebecca Henderson, Bengt Holmström, Michael Peters and Michele Tertilt for helpful comments as well as participants in seminars at the AEA, Barcelona, Berkeley, Birmingham, Bocconi, Brussels, CEU, Chicago, Dublin, Duke, Essex, George Washington, Harvard, Hong Kong, IIES, LBS, Leuven, LSE, Madrid, Mannheim, Michigan, Minnesota, MIT, Munich, Naples, NBER, NYU, Peterson, Princeton, Stockholm, Sussex, Toronto, Uppsala, USC, Yale and Zurich. The Economic and Social Research Council, the Kauffman Foundation, PEDL and the Alfred Sloan Foundation have given financial support. We received no funding from the global management consultancy firm (McKinsey) we worked with in developing the survey tool. Our partnership with Pedro Castro, Stephen Dorgan and John Dowdy has been particularly important in the development of the project. We are grateful to Daniela Scur and Renata Lemos for ongoing discussion and feedback on the paper. The views expressed herein are those of the authors and do not necessarily reflect the views of the National Bureau of Economic Research.

NBER working papers are circulated for discussion and comment purposes. They have not been peer-reviewed or been subject to the review by the NBER Board of Directors that accompanies official NBER publications.

© 2016 by Nicholas Bloom, Raffaella Sadun, and John Van Reenen. All rights reserved. Short sections of text, not to exceed two paragraphs, may be quoted without explicit permission provided that full credit, including © notice, is given to the source.

Management as a Technology?Nicholas Bloom, Raffaella Sadun, and John Van ReenenNBER Working Paper No. 22327June 2016JEL No. L2,M2,O32,O33

ABSTRACT

Are some management practices akin to a technology that can explain company and national productivity, or do they simply reflect contingent management styles? We collect data on core management practices from over 11,000 firms in 34 countries. We find large cross-country differences in the adoption of basic management practices, with the US having the highest size-weighted average management score. We present a formal model of “Management as a Technology”, and structurally estimate it using panel data to recover parameters including the depreciation rate and adjustment costs of managerial capital (both found to be larger than for tangible non-managerial capital). Our model also predicts (i) a positive effect of management on firm performance; (ii) a positive relationship between product market competition and average management quality (part of which stems from the larger covariance between management with firm size as competition strengthens); and (iii) a rise (fall) in the level (dispersion) of management with firm age. We find strong empirical support for all of these predictions in our data. Finally, building on our model, we find that differences in management practices account for about 30% of cross-country total factor productivity differences.

Nicholas BloomStanford UniversityDepartment of Economics579 Serra MallStanford, CA 94305-6072and NBERnbloom@stanford.edu

Raffaella SadunHarvard Business SchoolMorgan Hall 233Soldiers FieldBoston, MA 02163and NBERrsadun@hbs.edu

John Van ReenenDepartment of EconomicsLondon School of EconomicsCentre for Economic PerformanceHoughton StreetLondon WC2A 2AEUNITED KINGDOMand NBERj.vanreenen@lse.ac.uk

1 Introduction

Productivity differences between firms and between countries remain startling. For example, within

the average four-digit U.S. manufacturing industries, Syverson (2011) finds that labor productivity

for plants at the 90th percentile was four times as high as plants at the 10th percentile. Even after

controlling for other factors, Total Factor Productivity (TFP) was almost twice as high. These

differences persist over time and are robust to controlling for plant-specific prices in homogeneous

goods industries.1 Such TFP heterogeneity is evident in all other countries where data is available.2

One explanation is that these persistent within industry productivity differentials are due to “hard”

technological innovations, as embodied in patents or the adoption of advanced equipment. Another

explanation, which is the focus of this paper, is that productivity differences reflect variations in

management practices.

We advance the idea that some forms of management practices are like a “technology”, in the sense

that they raise TFP. This has a number of empirical implications that we examine and find support

for in the data. Our perspective on management is distinct from the dominant “Design” paradigm

in organizational economics, which views management as a question of optimal design depending

on the contingent features of a firm’s environment (Gibbons and Roberts, 2013). In this contingent

view of management practices, there is no sense in which any management styles are on average

better than any others. Our data provides some support for Design perspective, but we show that

- at least within the very stylized version of this perspective that we consider - this delivers only a

partial explanation of the patterns that we can observe in our data.

To date, empirical work to measure differences in management practices across firms and countries

has been limited. Despite this lack of data, the core theories in many fields such as international

trade, labor economics, industrial organization and macroeconomics are now incorporating firm

heterogeneity as a central component.3

To address the lack of management data, we collect original survey data on management practices

on over 11,000 firms in 34 countries. Besides its rich cross sectional nature, both in terms of

countries and industries covered, this dataset also features a significant panel component built

through four different survey waves from 2004 to 2014. We first present some stylized facts from

this database in the cross country and cross firm dimensions. One of the striking features of the

1These are revenue based measures of TFP (“TFPR”) so will also reflect firm-specific mark-ups. Foster, Halti-wanger and Syverson (2008) show large differences in TFP even within very homogeneous goods industries suchas cement and block ice. Hall and Jones (1999) and Jones and Romer (2010) show how the stark differences inproductivity across countries account for a substantial fraction of the differences in average income.

2Usually productivity dispersion is even greater than in other countries than in the U.S. - see Bartelsman, Halti-wanger and Scarpetta (2013) and Hsieh and Klenow (2009).

3Different fields have different labels for management. In trade, the focus is on an initial productivity draw whenthe plant enters an industry that persists over time (e.g. Melitz, 2003). In industrial organization, the focus hastraditionally been on cost heterogeneity due to entrepreneurial/managerial talent (e.g. Lucas, 1978). In macro,organizational capital is sometimes related to the firm specific managerial know-how built up over time (e.g. Prescottand Visscher 1980). In labor, there is a growing focus on how the wage distribution requires an understanding of theheterogeneity of firm productivity (e.g. Card, Heining and Kline, 2013).

2

data is that the average management score, just like TFP, is higher in the U.S. than it is in other

countries (see Figure 1). A second striking feature - shown in Figure 2 - is that management, like

TFP, shows a wide dispersion across firms within each country. Interestingly, this dispersion is

lower in countries like the U.S. with lower levels of market frictions than it is in countries like India

and Brazil.

We detail a simple model of “Management As a Technology” (MAT) which incorporates both

a heterogeneous initial draw of managerial ability when a firm starts up, and the endogenous

response of ongoing firms who change their level of managerial capital in response to shocks to

the environment (modeled as as idiosyncratic TFP shocks). The model is useful to formalize our

theoretical intuitions and enable structural estimation of key parameters. In particular, thanks to

the cross sectional and panel variation present in the management data, we are able to identify the

depreciation rate and adjustment costs of managerial capital, using Simulated Method of Moments

(SMM). A further benefit of the structural model is that it enables us to derive some additional

predictions on moments we did not target in the structural estimation.

We find that the data supports the predictions from the MAT model. First, management is posi-

tively associated with improved firm performance (e.g. productivity, profitability and survival), and

from experimental evidence the management effect appears to be causal. Second, firm management

rises with more intense product market competition, both through reallocating more economic ac-

tivity to the better managed firms (an Olley and Pakes,1996, covariance term), and also through a

higher unweighted average level of management. Third, older firms have a higher level of manage-

ment, but lower dispersion due to selection effects. We contrast our MAT model to the predictions

arising from a very stylized version of the “Management As Design” model, an alternative approach

which sees management as a contingent “Design” feature, rather than an output increasing factor of

production. There are some elements consistent with this second design approach, especially when

we disaggregate the management score into elements relating to monitoring compared relative to

incentives. However, overall the MAT model seems a better description of our data.

Finally, using our MAT model we show that on average just under a third of cross country TFP

differences with the U.S. are accounted for by management, with this fraction being higher in

OECD countries than in less developed nations. Thus, management practices can account for a

substantial portion of cross-country differences in development. Within this portion, 30% is due to

differences in the covariance between management and firm size - that is, differences in reallocation

effects - and the remaining 70% by differences in average unweighted management.

This paper relates to several literatures. First, there is a large body of empirical literature on the

importance of management for variations in firm and national productivity, going back to Walker

(1887) through to more recent papers like Ichniowski et al. (1997), Bertrand and Schoar (2003),

Adhvaryu, Kala and Nyshadham (2016) and Bruhn, Karlan and Schoar (2016). Second, there

is a growing macro literature on aggregate implications of firm management and organizational

structure, ranging from Lucas (1978), to Gennaioli et al (2013), Guner and Ventura (2014), Garicano

3

and Rossi-Hansberg (2015) and Akcigit, Alp and Peters (2016). Finally, there is another growing

literature focusing on explaining cross-country TFP in terms of the degree of reallocation out inputs

to more productivity firms, most notably Hsieh and Klenow (2009) and Restuccia and Rogerson

(2008).

The structure of the paper is as follows. We first describe some theories of management (Section

2) and how we collect the management data (Section 3). We then describe some of the data and

stylized facts (Section 4). Section 5 details our empirical results and Section 6 concludes.

2 Models of management

2.1 Conventional approaches to modeling heterogeneity in firm productivity

Econometricians have often labeled the fixed effect in panel data estimates of production functions

“management ability”. However, for the most part economists have focused on how technological

innovations drive economic growth; for example, correlating TFP with observable measures of

innovation such as R&D, patents, or information technology.

There is robust evidence of the impact of such “hard” technologies for productivity growth.4 Nev-

ertheless, there are at least two major problems in focusing on these aspects of technical change as

the sole cause of productivity dispersion. First, even after controlling for a wide range of observ-

able measures of hard technologies, a large residual in measured TFP still remains. Second, many

studies have found that the impact of technology on productivity varies widely across firms and

countries. In particular, information technology (IT) has much larger effects on the productivity

of firms that have complementary managerial structures which enable IT to be more efficiently

exploited.5Furthermore, a huge body of case study work also suggests a major role for management

in raising firm performance (e.g. Baker and Gil, 2013).

In light of these issues, we believe it is worth directly considering management practices as an

independent factor in raising productivity.

2.2 Formal models of management

It is useful to analytically distinguish between two broad approaches that we can embed in a simple

production function framework where value added, Y, is produced as follows:

Y = F (A, L,K,M) (1)

4For example, see Griliches (1998).5In their case study of IT in retail banking, for example, Autor et al (2002) found that banks who failed to

re-organize the physical and social relations within the workplace reaped little reward from new ICT (like ATMmachines). More generally, Bresnahan, Brynjolfsson and Hitt (2002) found that decentralized organizations tendedto enjoy a higher productivity pay-off from IT across a wide range of sectors. Similarly, Bloom, Sadun and VanReenen (2012) found that IT productivity was higher for firms with stronger incentives management (e.g. carefulhiring, merit based pay and promotion and vigorously fixing/firing under-performers).

4

where A is an efficiency term, labor is L, non-management capital is K, and M is management

capital.

We begin with the “Management as Technology” perspective, where some types of management

(or bundles of management practices) are better than others for firms across a wide range of

environments. There are three types of these “best practices”. First, there are some practices that

have always been better (e.g. not promoting incompetent employees to senior positions, or collecting

some information before making decisions). Second, there may be genuine managerial innovations

(e.g. Taylor’s Scientific Management; Lean Manufacturing; Deming’s Quality movement, etc.) in

the same way that there are technological innovations. Third, many management practices may

have become optimal due to changes in the economic environment over time. Incentive pay may be

an example of this, as the incidence of piece rates declined from the late 19th century, but appears

to be making a comeback today.6

The alternative model is the traditional approach in Organizational Economics, i.e. the “Man-

agement as a Design” perspective, where differences in practices are styles optimized to a firm’s

environment. For any indicator of M, such as the measures we gather, the Design approach would

not assume that output is monotonically increasing in M. In some circumstances, higher levels of

what we would regard as good practices will explicitly reduce output. To take a simple example,

consider M as a discrete variable which is equal to one if promotion takes into account effort and

ability and zero otherwise (e.g. purely seniority based promotions). The Design perspective could

find that pure tenure-based promotion, which ignores effort and ability, increases output in some

sectors, for example by reducing influencing activities (Milgrom, 1988), but increases it in others.

Under the Design approach, the production function can be written as equation (1), but for some

firms and practices F ′(M) ≤ 0. Even if M were costless, output would fall if it was exogenously

increased. The Design approach emphasizes that the reason for heterogeneity in the adoption

of different practices is that firms face different environments. This is in the same spirit as the

“contingency” paradigm in management science (Woodward, 1958).

The Design and the Technology perspectives can be nested within a common basic set-up but have,

as we show, very different theoretical and empirical implications. Leaving aside for the moment the

specific modeling choice of F (M), we formalize these ideas by treating M as an intangible capital

(as in Corrado and Hulten, 2010), which has a market price and also a cost of adjustment. We

allow firms to have an exogenous initial draw of M when they enter the economy. This creates ex

ante heterogeneity between firms (generalizing the approach in Hopenhayn, 1992, for TFP). Factor

inputs and outputs are firm specific (we do not use t subscripts for simplicity unless needed). We

consider a single industry, so firm-specific values are indicated by an i subscript

Yi = AiKαi L

βi G(Mi) (2)

6Lemieux et al (2009) suggest that this may be due to advances in IT. Software companies like SAP have madeit much easier to measure output in a timely and robust fashion, making effective incentive pay schemes easier todesign and implement.

5

where G(Mi) is a management function common to all firms. Demand is assumed to derive from a

final good sector (or equivalently a consumer) using a CES aggregator across individual inputs:

Y = N1

1−ρ

(N∑i=1

Yρ−1ρ

i

) ρρ−1

(3)

where ρ > 1 is the elasticity of substitution, N is the number of firms and N1

1−ρ is the standard

adjustment factor to make the degree of substitution scale free. Our main index of competition will

be ρ. Applying the first order conditions gives each firm an inverse demand curve with elasticity ρ

where we have normalized the industry price to be P = 1

Pi = (Y

N)1ρY−1ρ

i = BY− 1ρ

i

where the demand shifter is B = ( YN )1ρ. These production and demand curves generate the firm’s

revenue function:

PiYi = AiKai L

biG(Mi)

where for analytical tractability we defined Ai = Ai1−1/ρ

( YN )1ρ, a = α(1 − 1/ρ), b = β(1 − 1/ρ)

and G(Mi) = G(Mi)(1−1/ρ). Profits, defined as revenues less capital, labor and management costs

(cK(K), cL(L) and cM (M)), and fixed costs F are:7

Πi = AiKai L

biG(Mi)− cK(Ki)− cL(Li)− cM (Mi)− F

2.3 Models of management in production

In terms of the management function G(Mi), we consider two broad classes of models. First,

Management as a Technology where management is an intangible capital input in which output

is monotonically increasing. Second, Management as Design in which management is a choice of

production approach. We focus on the first as this fits the data substantially better (as we show

below) but lay out both approaches in what follows.

2.3.1 Management as a Technology (MAT)

In Lucas (1978) or Melitz (2003) style models, firm performance is increasing continuously in

the level of managerial quality, which is synonymous with productivity. Firms draw a level of

management quality when they are born, and this continues with them throughout their lives.

Since these types of models assume G(Mi) is increasing in Mi, we simplify the revenue function by

7Since firms in our data are typically small in relation to their input and output markets, for tractability we ignoreany general equilibrium effects, taking all input prices (for capital, labor and management) as constant.

6

assuming G(Mi) = M ci

PiYi = AiKai L

biM

ci

More generally, we want to allow for the possibility that management can also be endogenously

improved; for example, by hiring management consultants, spending time developing improved

organizational processes (e.g. Toyota’s Kaizen meetings), or paying for a better CEO. Although

managerial capital can be improved in this way, failure to invest may mean it depreciates over time

like other tangible and intangible assets such as physical capital, R&D, and advertising. Hence, we

set up a more general model which still has initial heterogeneous draws of management when firms

enter, but treats management as an intangible capital stock with depreciation:

Mit = (1− δM )Mit−1 + IMit IMit ≥ 0

where IMit reflects investment in management practices, which has a non-negativity constraint

reflecting the fact that managerial capital cannot be sold. The physical capital accumulation

equation is similar except it allows for capital resale with a resale loss of φK

Kit = (1− δK)Kit−1 + IKit − φKD[IKit < 0]

where D[IKit < 0] is an indicator function for negative investment (capital sales). Both manage-

rial and non-managerial investment goods can be purchased in the market at price wMt and wKt

respectively.

2.3.2 Management as Design

An alternative approach is to assume that management practices are contingent on a firm’s envi-

ronment, so that increases in M do not always increase output. In some sectors, high values of M

will increase output, and in others they will reduce output depending on the specific features of the

industry. We assume that optimal management practices may vary by industry and country, but

this could also occur across other characteristics like firm age, size, or growth rate. For example,

industries employing large numbers of highly skilled employees, like pharmaceuticals, will require

large investments in careful hiring, tying rewards to performance and monitoring output, while

low-tech industries can make do without these costly human resource practices. Likewise, optimal

management practices could vary by country if, for example, some cultures are more comfortable

with firing persistently under-performing employees (e.g. the U.S.) while others emphasize loyalty

to long-serving employees (e.g. Japan).

There are many ways to set up a Design model. As a simple example we define G(Mi) = 1/(1 +

θ|Mi−M |) where θ ≥ 0 and G(Mi) ∈ (0, 1] is decreasing in the absolute deviation of M from its opti-

mal levelM.8 There are of course many other ways to code this up - and this is certainly not meant to

8Our baseline case also assumes that M is a choice variable that does not have to be paid for on an ongoing basisso that δM = 0 although this assumption is not material.

7

represent the wide range of Design approaches - but is a simple example to illustrate the implications

of a G(Mi) function which attains an interior maximum (so that an optimal choice of management exists rather than higher values always increasing output).

2.3.3 Management as Capital?

We initially debated calling our main approach “Management as Capital” (rather than “Man-

agement as a Technology”), viewing management as an intangible capital stock (see for example

Bruhn, Karlan and Schoar (2010)). In the end, because of the evidence suggesting management

spillovers across plants within firms and between different firms (e.g. Greenstone, Hornbeck and

Moretti (2010), Atalay, Hortascu and Syverson (2014) and Braguinsky et al. (2015) and Bloom et

al. (2016)) we thought modeling management as a technology seemed more appropriate. However,

we recognize that either terminology could be used. Indeed, the classic technology input - the R&D

knowledge stock - is recorded as an intangible capital input by the Bureau of Economic Activity in

U.S. National Accounts.

2.4 Adjustment costs and dynamics

In general, changing a capital stock will mean bearing adjustment costs. This could reflect, for

example, the costs of the organizational resistance to new management practices (e.g. Cyert

and March, 1963 or Atkin et al. (2015)). We assume changing management practices involves

a quadratic adjustment cost:

CM (Mt,Mt−1) = γMMt−1(Mt −Mt−1

Mt−1− δM )2

where the cost is proportional to the squared change in management net of depreciation, and scaled

by lagged management to avoid firms outgrowing adjustment costs. This style of adjustment costs is

common for capital (e.g. Chirinko, 1993) and seems reasonable for management where incremental

changes in practices are likely to meet less resistance than large changes. Likewise, we also assume

similar quadratic adjustment costs for non-managerial capital:

CK(Kt,Kt−1) = γKKt−1(Kt −Kt−1

Kt−1− δK)2

To minimize on the number of state variables in the model, we assume labor is costlessly adjustable,

but requires a per period wage rate of w. Given this assumption on labor, we can define the optimal

choice of labor by ∂PY (A,K,L∗,M)∂L = w. Imposing this labor optimality condition and assuming the

MAT specification for management in the production function, we obtain:

Y ∗(A,K,M) = A∗Ka

1−bMc

1−b

8

where A∗ = bb

1−bA1

1−b and we normalize w to unity. Finally, ln(A) is assumed to follow a standard

AR(1) process so that ln(Ait) = lnA0+ρAln(Ai,t−1)+σAεi,t where εi,t ∼ N(0, 1). This will generate

the firm-specific dynamics in the model, which are an important feature of our data.

2.5 Optimization and equilibrium

Given the firm’s three state variables - business conditions A, capital K, and management M - we

can write a value function (dropping i -subscripts for brevity):

V (At,Kt,Mt) = max[V c(At,Kt,Mt), 0]

V c(At,Kt,Mt) = maxKt+1,Mt+1

[Y ∗t − CK(Kt+1,Kt)− CM (Mt+1,Mt)− F

+ βEtV (At+1,Kt+1,Mt+1)]

where the first maximum reflects the decision to continue in operation or exit (where exit occurs

when V c < 0), the second (V c for “continuers”) is the optimization of capital and management

conditional on operation, and β is the discount factor. We assume there is a continuum of potential

new entrants that would have to pay an entry cost κ to enter. Upon entry, they take a stochastic

draw of their productivity and management values from a known joint distribution H(A,M) and

start with K0 = 0. Hence, entry occurs until the point that

κ =

∫V (A,K0,M)dH(A,M)

We solve for the steady-state equilibrium by selecting the demand shifter (B = ( YN )1ρ) that ensures

that the expected cost of entry equals the expected value of entry given the optimal capital and

management decisions. This equilibrium is characterized by a distribution of firms in terms of their

state values A,K,M . The distribution of lnA is assumed normal, while M is assumed to be drawn

from a uniform distribution.9

2.6 Numerical Estimation

Solving the model requires finding two nested fixed-points.10 First, we solve for the value functions

for incumbent firms using the contraction mapping (e.g. Stokey and Lucas, 1986), taking demand

as given for each firm. The policy correspondences for M and K are formed from the optimal

choices given these value functions, and for L from the static first-order condition. Second, we then

iterate over the demand curve (3) to satisfy the zero-profit condition.11 Once both fixed points

9Nothing fundamental hinges on the exact distributional assumptions for M and A.10The full replication package for the simulation and SMM estimation is available on

http://web.stanford.edu/˜nbloom/MAT.zip

11If there is positive expected profit then net entry occurs and the demand shifter B = ( YN

)1ρ

falls, and if there isnegative expected profit then net exit occurs.

9

are satisfied, we simulate data for 5,000 firms over 100 years to get to an ergodic steady-state, and

then discard the first 90 periods to keep the last 10 years of data (to match the time span of our

management panel data).

To solve and simulate this model we also need to define a set of 15 parameter values. We pre-define

nine of these from from the prior literature, normalize two (fixed costs to 100 and the mean of

ln(TFP) to 1) and estimate the remaining four parameters on our management and accounting

data panel. The nine predefined parameters are listed in Table 1, and are all based on standard

values in the literature. The four estimated parameters are those where much less is known from

the literature. The adjustment cost (γM ) and depreciation rates (δK) for management have never

been estimated before, to our knowledge. The sunk cost of entry (κ) is also hard to know. Finally,

we also estimate the adjustment cost for non-managerial capital (γK).12

To estimate the model by SMM we picked four data moments to match: the exit rate to help

inform the sunk cost entry, and the variance of the five-year growth rates of the three state vari-

ables (management capital, non-management capital, and TFP) to tie down the adjustment cost

and depreciation parameters. These data moments were generated on the matched management-

accounting panel dataset for all countries from 2004 to 2014 (described in more detail in the next

section). To generate standard-errors, we block-bootstrapped over firms the entire process 1,000

times to generate the variance-covariance matrix, which was also used to optimally weight the SMM

criterion function (see Appendix C for details).

2.7 Simulation results

The top panel of Table 2 contains the SMM estimates and standard errors values for the four

estimated parameters, and the bottom panel contains the moments from the data used to estimate

these. Because the model is exactly identified we can precisely match the moments within numerical

rounding errors.

The estimation of the adjustment costs for management is one of the novel contributions of this

paper. We obtain a slightly higher level of adjustment costs for management of 0.207 (compared to

0.189 for capital) which, alongside the irreversibility of management, helps generate smoother man-

agement five-year growth moments compared to capital five-year growth moments (see the bottom

panel of Table 2).13 These magnitudes are prima facie plausible - economic intuition (Cyert and

March, 1963) and anecdotal evidence from the private equity and management consulting industry

suggest that management practices are likely to be harder to change than plant or equipment.

Depreciation of management capital is is 13.3%, similar to the level of the depreciation of capital

12While prior papers have estimated labor and capital adjustment costs (e.g. Bloom, 2009, and the survey therein)they ignore management as an input so it is not clear these parameters are transferable to our set-up.

13If we allow management to be have the same 50% resale loss as capital it’s adjustment cost is estimated to be0.290, about 50% higher than the value for capital.

10

(10% - see Table 1).14 Finally, we obtain a sunk cost of entry that is 166% of the ongoing annual

fixed cost of running a plant.

Having defined and estimated the main MAT model, we can proceed to examine covariances of

various moments that we have not targeted in the structural estimation to later compare these

with actual data. Figures 3 through 5 show some predictions arising from the simulation. In

Figure 3, we start by comparing the distribution of management practices of a random draw of

15,489 firm-years from our simulation to the 15,489 firm-year surveys in the management panel

data, revealing similar cross-sectional distributions.15 While this is not a formal test of our model,

it does confirm it can generate the wide spread of management practices that is a striking finding

of the management survey data. Figures A2 and A3 show the unsurprising result that firm size

and TFP are increasing in the firm-level value of management.

Figure 4 examines the relationship between management and product market competition as in-

dexed by the elasticity of demand (ρ). We run all the simulation for increasingly high levels of the

absolute price elasticity of demand between three and fifteen (recall that our baseline is an elasticity

is equal to five). This represents economies with increasingly high levels of competition. We see

that average management scores are higher when competition is stronger. The darker bars are the

unweighted means of management across firms - they rise because under higher competition poorly

managed firms tend to exit as they cannot cover their fixed cost of production. We also see that size

(employment) weighted management practices rise even faster with competition because this raises

the covariance between firm size and management (a higher “Olley Pakes reallocation” term), as

better managed firms will acquire larger market shares (and therefore need more inputs). Finally,

Figure 5 examines the relationship between management and firm age. Firms’ management score

rises with age as poorly managed firms either improve or exit the market. Over time this leads the

dispersion of management practices to fall within any age cohort, because of the contraction of the

left tail of poorly managed firms.

Figure 6 Panels A to C provide similar figures to Figures 3 to 5 for our Management as Design

model, in which we assume G(M) is maximized at M = 3 for illustrative purposes. In Panel

A, we see a similar spread of management practices, suggesting the Design view can generate an

equilibrium dispersion of management practices. But in Panel B we see a very different relationship

with competition, where management practices are invariant with the level of competition. More

specifically, there is no sense in which high levels of management are better, and therefore they

are not positively selected as competition increases. In Panel C, we also see no variation in the

average management score with age for similar reasons, although we do see some reduction in

variance with age as extremely high and low values of management practices are modified or the

firm exits. Finally, in Panel D we have also included a plot of performance in terms of sales against

14One interpretation is that management capital is tied to the the identity of plant managers. The average jobtenure for plant managers in our survey is 6.4 year in the post and and 13.0 years in the company, which would implypost and company quit rates of about 15% to 7% spanning the depreciation estimate of 10%.

15To scale our management practices we take logs of the management variable, and normalize the lowest value to1 and the higher value to 5 to replicate our management survey scoring tool.

11

management, showing the inverted U shape implied by the Design view of the world that firms

have an optimizing level of M at 3.

3 Data

3.1 WMS Survey method

We describe the datasets in more detail in Appendix A, but sketch out the important features

here. To measure management practices, we developed a survey methodology known as the World

Management Survey (WMS).16 This uses an interview-based evaluation tool that defines 18 basic

management practices and scores them from one (“worst practice”) to five (“best practice”) on a

scoring grid. This evaluation tool was first developed by an international consulting firm, and scores

these practices in three broad areas.17 First, Monitoring : how well do companies track what goes on

inside their firms, and use this for continuous improvement? Second, Target setting : do companies

set the right targets, track outcomes, and take appropriate action if the two are inconsistent?

Third, Incentives/people management18: are companies promoting and rewarding employees based

on performance, and systematically trying to hire and retain their best employees?

To obtain accurate responses from firms, we interview production plant managers using a “double-

blind” technique. One part of this technique is that managers are not told in advance they are

being scored or shown the scoring grid. They are only told they are being “interviewed about

management practices for a piece of work”. The other side of the double blind technique is that

the interviewers do not know anything about the performance of the firm.

To run this blind scoring, we used “open” questions. For example, on the first monitoring question

we start by asking the open question, “tell me how your monitor your production process”, rather

than closed questions such as “Do you monitor your production daily? [yes/no]”. We continue

with open questions focused on actual practices and examples until the interviewer can make an

accurate assessment of the firm’s practices. For example, the second question on that performance

tracking dimension is, “What kinds of measures would you use to track performance?” and the

third is “If I walked around your factory, could I tell how each person was performing?”.19

The other side of the double-blind technique is that interviewers are not told anything about the

firm’s performance in advance. They are only provided with the company name, telephone number,

16More details can be found at http://worldmanagementsurvey.org/17Bertrand and Schoar (2003) focus on the characteristics and style of the CEO and CFO, and more specifically

on differences in strategic management (e.g. decision making applied to mergers and acquisitions), while Lazear,Shaw and Stanton (2016) focus on individual supervisors. The type of practices we analyze in this paper are closerto operational and human resource practices.

18These practices are similar to those emphasized in earlier work on management practices, by for example Ich-niowski, Prennushi and Shaw (1997) .

19The full list of questions for the grid is in Table A1 and (with more examples) athttp://worldmanagementsurvey.org/wp-content/images/2010/09/Manufacturing-Survey-Instrument.pdf.

12

and industry. Since we randomly sample medium-sized manufacturing firms (employing between

50 and 5,000 workers) who are not usually reported in the business press, the interviewers will

generally have not heard of these firms before, so they should have few preconceptions.20

The survey was targeted at plant managers, who are senior enough to have an overview of manage-

ment practices but not so senior as to be detached from day-to-day operations. We also collected

a series of “noise controls” on the interview process itself - such as the time of day, day of the

week, characteristics of the interviewee, and the identity of the interviewer. Including these in our

regression analysis typically helps to improve our estimation precision by stripping out some of the

random measurement error.

To ensure high sample response rates and informative interviews, we hired students with some

business experience and training. We also obtained government endorsements for the surveys in

each country covered. We also never asked interviewees for financial data, obtaining this instead

from independent sources on company accounts.

Finally, the interviewers were encouraged to be persistent - so they ran about two interviews a

day lasting 45 minutes each on average, with the rest of the time (about 6 hours a day) spent

repeatedly contacting managers to schedule interviews. This process, while time consuming and

expensive, helped to yield a 41% response rate which was uncorrelated with the (independently

collected) performance measures.

3.2 Survey waves

We have administered the survey in several waves since 2004. There were five major waves in

2004, 2006, 2009/10, 2013, and 2014. In 2004 we surveyed four countries (France, Germany, the

U.K. and the U.S.). In 2006 we expanded this to twelve countries (including Brazil, China, India,

and Japan), continuing random sampling, but in addition to a refreshment sample for the 2004

countries we also re-contacted all of the original 2004 firms to establish a panel. In 2009/10 we

re-contacted all the firms surveyed in 2004 and 2006, but did not do a refreshment sample (due to

budgetary constraints). In 2013 we added an additional number of countries (mainly in Africa and

Latin America). In 2014 we again did a refreshment sample, but also followed up the panel firms

in the U.S. and some E.U. countries. The final sample includes 34 countries and a panel of up to

four different years between 2004 and 2014 for some firms. In the full dataset we have 11,383 firms

and 15,489 interviews where we have usable management information.

20We focus on firms over a size threshold because the formal management practices we consider are likely to beless important for smaller firms. We had a maximum size threshold because we only interviewed one or two plantmanagers in each firm, so would have too incomplete a picture for very large firms. Below, we show tests suggestingour results are not biased by using this sampling scheme (see Appendix B).

13

3.3 Internal validation

We re-surveyed 5% of the sample using a second interviewer to independently survey a second

plant manager in the same firm. The idea is that the two independent management interviews

on different plants within the same firms reveal how consistently we are measuring management

practices. We found that in the sample of 222 re-rater interviews, the correlation between our

independently run first and second interview scores was 0.51 (p-value 0.001). Part of this difference

across plants within the same firm is likely to be real internal variations in management practices,

with the rest presumably reflecting survey measurement error. The highly significant correlation

across the two interviews suggests that while our management score is clearly noisy, it is picking

up significant management differences across firms.

3.4 Some descriptive statistics

The bar chart in Figure 1 plots the average (unweighted) management practice score across coun-

tries. This shows that the U.S. has the highest average management practice score, with the

Germans, Japanese, Swedes, and Canadians below, followed by a block of West European countries

(e.g. France, Italy and the U.K.) and Australia. Below this group is Southern European countries

(e.g. Portugal and Greece) and Poland. Emerging economies (e.g. Brazil, China, and India) are

next, and low income countries (mainly in Africa) are at the bottom. In one sense this cross-country

ranking is not surprising since it approximates the cross-country productivity ranking. But the cor-

relation is far from perfect - Southern European countries do a lot worse than expected and other

nations, like Poland and Mexico, do better.21

A key question is whether management practices are uniformly better in some countries like the

U.S. compared to India, or if differences in the shape of the distribution drive the averages? Figure

2 plots the firm-level histogram of management practices (solid bars) for all countries pooled (top

left) and then for each country individually. This shows that management practices, just like firm-

level productivity, display tremendous variation within countries. Of the total firm-level variation

in management only 13% is explained by country of firm location, a further 10% by industry

(measured at the three digit SIC level), with the remaining 77% being within country and industry.

Interestingly, countries like Brazil and India have a far larger left tail (e.g. scores of two or less) of

badly run firms than the U.S. .22 This immediately suggests that one reason for the better average

performance in the U.S. is that the American economy is better at selecting out the badly managed

firms. We pursue the idea that the U.S. advantage may be linked to stronger forces of competition

below.

21Polish management appears to be better because of the influence of the large numbers of German multinationalsubsidiaries, while Mexico similarly benefits from a heavy U.S. multinational presence

22For example, the skewness of the firm level management distribution in the U.S. is 0.09, whereas the skewness ofthe distribution in Brazil is 0.16 and 0.36 in India.

14

Figure A1 shows average management scores in domestic firms (i.e. those who are not part of

groups with overseas plants) compared to plants belonging to foreign subsidiaries. The average

scores in domestic plants look similar to those in Figure 1, which is unsurprising as most of our

firms are domestic. More interesting is that plants belonging to foreign multinationals appear

to score highly in almost every country, suggesting that such firms are able to transplant their

management practices internationally. This finding - which is robust to controlling for many other

factors (such as firm size, age and industry) - is consistent with the idea of a subset of global,

productivity enhancing practices. An interesting extension to our basic model would be to allow

for this type of cross-plant transfer of management practices (e.g. Helpman, Melitz and Yeaple,

2004) but for parsimony in the current model we have not done so.

3.5 Managerial and Organizational Practices Survey (MOPS)

We also implemented a more traditional closed question “tick box” survey design for MOPS which

gives us management data on 31,793 U.S. manufacturing plants in 2010. The question design was

modeled on WMS and the response to the MOPS was very high as we worked with the U.S and

replies were legally mandatory. Census Bureau. Details on MOPS is in Bloom et al (2016) and

Appendix A. One advantage of MOPS is that it has much more reliable information on plant and

firm age than in WMS - as discussed in later sections of this paper - so we use MOPS for one of

our theoretical predictions on the relationship between management and age.

4 Implications of Management as a Technology

4.1 Management and firm performance

Basic results

The most obvious implication of the MAT model is that high management scores should be asso-

ciated with better firm performance. Figure A2 plots firm sales on firm management and Figure

A3 does the same for conventionally measured firm TFP and management scores using local linear

regressions. Both figures show a clear positive and monotonic relationship. To probe this bivariate

relationship more formally, we run some simple regressions. We z-score each individual practice,

average across all 18 questions, and z-scored this average so the management index has a stan-

dard deviation of unity.23 Table 3 examines the correlation between different measures of firm

performance and management. To measure firm performance we used company accounts data24,

23We have experimented with other ways of aggregating the management scores such as using principal componentanalysis. Since the 18 questions are all positively correlated these more sophisticated alternatives produce broadlysimilar results to those developed here. Sub-section 4.5 below describes some other ways of dis-aggregating the scoresinto sub-components that reveals evidence for the Design perspective.

24Our sampling frame contained 90% private firms and 10% publicly listed firms. In most OECD countries bothpublic and private firms publish basic accounts. In the U.S., Canada and India, however, private firms do not publish

15

estimating production functions where Qit is proxied by the real sales of firm i at time t :

lnQit = αMMit + αLlnLit + αK lnKit + αXxit + uit (4)

where Mit is the empirical management score25, xit is a vector of other controls such as the pro-

portion of employees with a college degree, noise controls (e.g. interviewer dummies), country and

three digit SIC industry dummies and uit is an error term. In column (1) of Table 3 we regress

ln(sales) against ln(employment) and the management score, finding a highly significant coefficient

of 0.356. This suggests that firms with one standard deviation of the management score are associ-

ated with 36 log points higher labor productivity (i.e. about 43%). In column (2) we add the capital

stock and other controls which causes the coefficient on management to drop to 0.159, although it

remains significant. Column (3) conditions on a sub-sample where we observe each firm in at least

two years to show the effects are stable, while column (4) re-estimates the specification including

a full set of firm fixed effects to identify from changes in management over time, a very tough test

given the likelihood of attenuation bias. The coefficient on management (and labor and capital)

does fall, but remains positive and significant.26 In column (5) we instead use the Olley and Pakes

(1996) estimator of productivity and obtain a significant management coefficient of 0.231.

As discussed above, one of the most basic predictions is that better managed firms should be

larger than poorly managed firms. Column (6) of Table 3 shows that better managed firms are

significantly larger than poorly managed firms with a one standard deviation of management asso-

ciated with a 40 log point (49%) increase in employment size. In column (7) we use profitability

as the dependent variable as measured by ROCE (Return on Capital Employed) and show again

a positive association with management. Considering more dynamic measures, column (8) uses

sales growth as a dependent variable, revealing that better managed firms are significantly more

likely to grow. Column (9) estimates a model with Tobin’s average q as the dependent variable,

which is a forward looking measure of performance. Although this can only be implemented for

the publicly listed firms, we see again a positive and significant association with this stock market

based measure. Finally, column (10) examines bankruptcy/death and finds that better managed

firms are significantly less likely to die.

These are conditional correlations that are consistent with the MAT model, but are obviously not

to be taken as causal. However, the randomized control trial (RCT) evidence in Indian textile

firms (Bloom et al, 2013) showed that increasing WMS style management scores by one standard

deviation in management caused a 10% increase in TFP. This estimate lies between the fixed effect

estimates of column (4) and the cross sectional estimates of column (3). Other well identified

(sufficiently detailed) accounts so no performance data is available. Hence, these performance regressions use datafor all firms except privately held ones in the U.S., Canada and India.

25The empirical measure of management here, M, corresponds to the log of the managerial capital stock (lnM ) inthe theory. This seems reasonable given the evidence of Figure 3 of the log-normal distribution of the empirical score.

26Note that these correlations are not simply driven by the “Anglo-Saxon” countries, as one might suspect if themanagement measures were culturally biased. We cannot reject that the coefficient on management is the same acrossall countries: the F-test (p-value) on the inclusion of a full set of management*country dummies is 0.790 (0.642).

16

estimates of the causal impact of management practices - such as the RCT evidence from Mexico

discussed in Bruhn, Karlan and Schoar (2016) and the management assistance natural experiment

from the Marshall plan discussed in Giorcelli (2016) - find similarly large impacts of management

practices on firm productivity.

4.2 Product Market Competition

4.2.1 Competition and management

An important implication of the management as technology model is that tougher competition is

likely to improve average management scores. To test this prediction, we estimate regressions of

the form:

Micjt = γCOMPETITIONcjt + αzit + ηt + ξcj + νicjt (5)

where zit is a vector of other firm controls (the proportion of employees with a college degree, log

firm and plant size, log firm age and noise controls), ηt denotes year dummies, ξct denotes a full set

of three digit SIC industry dummies by country, and v is an error term.

We employ three different industry measures of competition. First, we begin with the inverse

industry Lerner index measured in an industry by country by period cell. The Lerner index is

a classic measure of competition (Aghion et al, 2005), and is calculated as the median price cost

margin within an industry-country cell using all firms in the ORBIS accounting database.27 Since

profits data is not generally reported for firms in developing countries, we focus on OECD countries.

We build a time varying Lerner index using data relative to three different periods (2003-2006; 2008-

2011; 2012-2013).28 These industry by country by period variables are then correlated with the

management scores conducted over the same time periods.

As an alternative to the Lerner measure of competition, we use a measure of import penetration

(imports over apparent consumption) in the country by industry by period cell, again measured

in the same periods and for the same set of OECD countries using industry by country by year

data from the World Input-Output Database (WIOD). Finally, to take into account the fact that

observed changes in import competition may not be exogenous, we build an alternative measure of

import penetration from WIOD which includes only imports from China, as these have been show

in other papers (e.g. Bloom, Draca and Van Reenen, 2015) to be overwhelmingly driven by policy

changes such as Chinese accession to the WTO and the subsequent reduction in tariffs and quotas

(e.g. the dismantling of the Multi-Fiber Agreement).

27In the simulated data we confirm that this empirical measure of the Lerner Index is highly correlated with ourconsumer price sensitivity parameter, ρ. For example, the Lerner has a correlation of 0.928 with price sensitivityacross simulations in which we increase ρ in unit increments from 3 to 15.

28See the Appendix for details on the construction of the measures of competition. These roughly correspond toblocks of time before, during and after the Great Recession/Euro Crisis. 2013 is the last full year of the ORBISdatabase currently available.

17

We begin by just showing the raw data in Figure 7, binning the three competition measures into

terciles and plotting the mean management score in each bin. Panel A shows this for cross sec-

tional “levels” (after subtracting the overall industry means and overall country means in both the

competition measure and the management score), revealing a robustly positively relationship for all

three competition measures. Panel B reports a similar graphic for “changes” in management over

time within a country by industry pair (i.e. subtracting the country by industry means) against

changes in competition over time, again displaying a robustly positive relationship.

In Table 4, we examine this more formally in a regression context estimating equation (5). The

dependent variable across all columns is the standardized value of the management score. Column

(1) reports the correlation between management and competition including industry by country

fixed effects, time dummies and other standard firm-level controls. Consistent with Figure 7, the

Lerner Index has a positive and significant correlation with management. The simulation model

suggests that this relationship should be stronger if we size-weight management due to better

reallocation in more competitive sectors. Column (2) does this using as a weight the share of

employment in the industry by country cell. Indeed, the coefficient on the Lerner measure rises from

0.99 to 1.75. The next four columns repeat the specifications but use import penetration, including

imports from all countries in columns (3) and (4) and just imports from China in columns (5) and

(6), as an alternative measure of competition. The pattern of results shows a larger coefficient on

the competition measures for the size-weighted regressions compared to the unweighted regressions,

consistent with the findings from using the Lerner Index.

We also considered a fourth measure of competition from our survey data: the number of rivals

as perceived by the plant manager. The advantage of this indicator is that it is available for all

countries in our survey. Empirically, the variable is also linked to improvements in management. In

a specification like column (1) the coefficient (standard error) on this measure of competition was

0.033 (0.017) on a sample of 14,305 observations including all countries with management data,

and 0.059 (0.022) on the sample of OECD countries overlapping with the one used in Table 4

(8,414 observations as there are a few some missing values on the number of competitors variable).

The disadvantage of the number of rivals measure is that it is not tightly linked to the theory

simulations.29

Overall, the results suggest that higher competition is associated with significant improvements in

management, and the magnitude of the coefficient is larger when we weight the regressions by size.

In terms of magnitudes a one standard deviation change in the Lerner index in the unweighted

regression is associated with a 0.06 of a standard deviation change in management, compared to

0.02 using the import penetration measure and 0.05 using Chinese imports. The equivalent numbers

for the weighted regressions are 0.11, 0.05 and 0.05.30

29Although falls in barriers to entry will tend to increase the number of firms in the MAT model, increases inconsumer sensitivity to price can lead to an equilibrium reduction in the number of firms.

30To check whether the difference between the weighted and the unweighted results was significant, we comparedthe distribution of the estimated coefficients with and without weights bootstrapping with 500 replications. The

18

4.2.2 Competition and reallocation towards better managed firms

Another way to confirm the reallocative impact of competition predicted by the MAT model is to

consider how factors that reduce the degree of competition reduce the covariance between manage-

ment practices and firm size, implying γ < 0 in the following equation:

FirmSizeit = γ (M ∗ COMPETITION)it + δ1Mit + δ2COMPETITION i + δ3xijt + νijt (6)

The simplest method of testing this idea is to use countries grouped into regions to proxy competi-

tive frictions, as it is likely that competition is stronger in some regions (e.g. the U.S.) than others

(e.g. southern Europe).

Column (1) of Table 5 reports the results of a regression of firm employment on the average

management score and a set of industry, year and country dummies.31 The results indicate that

increasing a firm’s management score by one standard deviation is associated with an extra 183

workers. In column (2) we allow the management coefficient to vary by region, with the U.S. as the

omitted base. The negative coefficients on the interactions indicate that that there is a stronger

relationship between size and management in the U.S. compared to other regions. This difference

is significant for Africa, Latin America and southern Europe, but not for Asia or northern Europe.

A one standard deviation increase in management is associated with 268 extra employees in the

U.S. but only 68 (= 268.4 - 199.5) extra workers in southern Europe.32 These results suggest that

reallocation is stronger in the U.S. than in the other countries, which is consistent with the findings

on productivity in Bartelsman, Haltiwanger and Scarpetta (2013) and Hsieh and Klenow (2009).

We also investigate explicit measures of market-friction variables that can reduce competition. In

columns (3) to (5) of Table 5 we use country-wide measures of employment regulation from the

OECD and trade costs from the World Bank. Both of these reduce the covariance between firm size

and management practices. Finally in column (6) we use the more detailed country by industry

measures of tariffs from Feenstra and Romalis (2012) in deviations from their country and industry

mean, and again find a significant negative interaction. This implies that within a sector (like

steel), countries with higher tariffs (such as Brazil) will allocate less activity to better managed

firms than those with lower tariffs (such as the U.K.).

weighted coefficients were larger than the unweighted coefficients 84% of the times for the Lerner index, 76% of thetimes with the import penetration variable, and 52% of the times using imports from China.

31This is the measure of firm size reported by the plant manager. For a multinational this may be ambiguous as theplant manager may report the global multinational size which is not necessarily closely related to the managementpractices of the plant we survey. Consequently, Table 5 drops multinationals and their subsidiaries, but we showrobustness of this procedure below.

32These results are for covariances based on size. Using a dynamic version of this moment - the covariance betweenemployment growth and management - generates qualitatively similar results. For example, re-running column (2)using the growth (rather than the level) of employment also has negative interactions on all the regional interactions.For example, a one standard deviation increase in management in the U.S. raises sales growth by 6.9% compared toa (significantly lower) 0.5% faster growth in Asia from a similar increase in management.

19

Taken as a whole, these findings on competition appear very consistent with the predictions of our

MAT model.

4.3 Age

Examining the third prediction from MAT - the relationship between firm age and management -

is complicated by the fact that the “date of incorporation” information in company accounts refers

to the year in which the company was formed, even if this is due to a merger or acquisition.33

Consequently, we turn to a complementary management database, the Management and Organiza-

tional Practices (MOPS) survey, which has more accurate age data based on plant age rather than

firm age.34 MOPS is a plant-level survey with management questions that we helped design with

very similar questions to those in our standard telephone survey. Figure 8 shows strong evidence

that the average management score is higher in older cohorts of plants, and that the variance of

management scores is lower. This relationship is particularly strong when comparing plants who

have been in existence for five or less years with their older counterparts. This closely matches the

predictions from the simulation model, in which the exit of establishments with low management

draws after birth increases the average management score and reduces the management variation

(see Figure 4).35

4.4 Other predictions from the MAT model - the price of management

There are other rich predictions from MAT. One obvious implication is that managerial capital

should fall as its price increases. But how can we measure this price? It is plausible that the supply

of highly educated workers is a complement to managerial ability, especially from institutions

that supply managerial education (e.g. Gennaioli et al, 2013). To examine this idea, we used

GIS software to geocode the latitude and longitude of every plant in our database and performed a

similar exercise for every college and business school using the UNESCO Higher Education Database

(Feng, 2013) which records he location of every university and business school in the world down

to the zipcode level. We then calculated the driving times to the nearest university/business school

for each of our plants.

The WMS management score significantly increases the closer the plant is to a leading university

or business school (see Table A6). This is true even controlling for population density, regional

33For example, a company like GSK is denoted as formed in 2001 when Glaxo Wellcome merged with Smithkline-Beecham, even though Glaxo-Wellcome has a history dating back to late Nineteenth Century (Jason Nathan andCompany, started in 1873, merged with Burroughs Wellcome and Company, started by Henry Wellcome and SilasBurroughs in 1880).

34Plant age in the Census is measured from the first year of existence in the Census/IRS Business Registry, whichis built from social security and income tax records.

35MOPS was also linked to productivity data in the Annual Survey of Manufacturers and Census of Manufactur-ing. (Bloom et al, 2016) show that we obtain similar results on the positive connection between higher plant levelmanagement scores and performance as shown in Table 3 above, and the positive correlation of management withcompetition as shown in Table 4 above.

20

dummies, weather conditions, distance to the coast, and a host of other variables. The proportion

of more educated employees and managers also significantly increases with proximity to a university

(as one would expect if there are mobility frictions and graduates are more likely to find employment

in a nearby firm). So a plausible reason for the positive correlation of management with universities

is through the supply of skills, reducing the cost of investing in managerial capital.

4.5 Management as Design

The predictions of the Management as a Technology model on performance, competition, age and

price are all consistent with the results from the WMS and MOPS management datasets. Our

extremely stylized version of the Management as Design model does less well. This Design model

does successfully predict the dispersion of management (compare Figure 3 with Figure 6 Panel

A) and the falling dispersion of management with age (compare Figure 8 light bars with Figure 6

Panel C). However, the predictions of a non-monotonic relationship between firm performance and

management are rejected (compare Figure A2 with Figure 6 Panel D) as is the flat relationship

between management and competition (compare Figure 7 with Figure 6 Panel B) and management

and age (compare Figure 8 dark bars with Figure 6 Panel C).

One set of results that is instead consistent with the Design approach relates to the contingency of

specific types of management practices on different industry characteristics. More specifically, the

Design approach suggests we might expect sectors that make intensive use of tangible fixed capital to

specialize more in monitoring/targets management, whereas human capital intensive sectors focus

more on people/incentives management. This is indeed what we tend to observe when we correlate

our management data with four digit U.S. industry data on the capital-labor ratio (NBER) and

R&D per employee (NSF), as shown in Panel A of Table 6.36 Although both people management

(column 1) and monitoring/targets management (column 2) are increasing in capital intensity, the

relationship is much stronger for monitoring & targets, as shown when we regress the relative

variable (people/incentives score minus monitoring/targets score) on capital intensity in column

(3). The opposite is true for R&D intensity as shown in the next three columns: in high tech

industries, people management is much more important. These findings are robust to including

them together with skills in the final three columns. As an alternative empirical strategy in Panel

B, we use country by industry specific values of these variables from the EU-KLEMS dataset. In

these specifications we are using the country-specific variation in capital and R&D intensity within

the same industry. The results are qualitatively similar to Panel A - capital intensive industries

have higher monitoring/target management practices, while R&D intensive industries have higher

people management practices scores, consistent with a basic Design model.

In summary, MAT appears to provide the best all around fit for the data, particularly in terms

of firm performance. We will use the implications of this model in the next section to calculate

36This is implicitly assuming that the U.S. values are picking up underlying technological differences betweenindustries that are true across countries.

21

what share of cross-country differences in TFP can be attributed to differences in management

practices. However, there is some support for the Design model in terms of contingent management

styles, suggesting that a hybrid model could offer a slightly better fit but at the expense of greater

complexity.

5 Accounting for cross-country TFP differences with Manage-

ment

We turn to a long-standing question in economics, stretching back to at least Walker (1887), of

how much of the variation in national and firm performance can be accounted for by differences in

management practices? We begin at the country level by defining an aggregate country management

index and decomposing this into a within firm and between firm component following Olley and

Pakes (1996) as:

M =∑i

Misi =∑i

[(Mi −Mi

)(si − si)

]+Mi = OP +Mi

where Mi is the management score for firm i, si is a size-weight (the firm’s share of employment

in the country), M is the unweighted average management score across firms and OP indicates the

“Olley Pakes” covariance term,∑

i

[(Mi −Mi

)(si − si)

]. The OP term simply divides manage-

ment into a within and a between/reallocation term. Comparing any two countries k and k’, the

difference in weighted scores can be decomposed into the difference in reallocation and unweighted

management scores:

Mk −Mk′ =(OP k −OP k′

)+(Mi

k −Mk′

i

)A deficit in aggregate management is composed of a difference in the reallocation effect

(OP k −OP k′

)and the average unweighted firm management scores

(Mi

k −Mk′

i

). Note that one could replace

Management by TFP or labor productivity for a more conventional analysis.

Table 7 reports the results of this decomposition (more details in Appendix B) using the U.S. as

the base country as it has the highest management scores. In column (1) we present the employ-

ment share-weighted management scores (M) in z-scores, so all differences can be read in standard

deviations. These differ from those presented earlier in Figure 1 because we have dropped multina-

tionals (to focus on clean national differences) and size-weighted the management scores. In column

(2) we show the unweighted average management score (Mi), and in column (3) the Olley Pakes

covariance term. From this we can see that, for example, the leading country - the U.S. - has a

size-weighted management score of 0.90, which is split almost half in between a reallocation effect

(0.40) and an unweighted average management score effect (0.50). Thus, the U.S. not only has

the highest unweighted management score but it also has a high degree of reallocation as discussed

22

above in sub-section 4.2.37

We next calculate each country’s management gap with the U.S. Column (4) does this for the

overall management gap and column (5) reports the share of this gap arising from differences in

reallocation. These results are also presented graphically in Figure 8, which shows that reallocation

accounts for a non-trivial fraction of the management gap in just about all countries.

We can push this analysis further by examining how much management could explain cross country

differences in TFP. Column (6) of Table 7 contains the country’s TFP gap with the U.S. using the

latest Penn World Tables (Feenstra, Inklaar and Timmer, 2015).38 Following the randomized

control trial (Bloom et al, 2013) and non-experimental evidence in Table 3, we assume that a one

standard deviation increase in management causes a 10% increase in TFP. For example, we can

estimate that improving Greece’s weighted average management score to that of the U.S. (a 1.3

standard deviation increase) would increase Greek TFP by 13%, about a third of the 37.5% TFP

gap between Greece and the U.S. Column (7) contains similar calculations for the other countries.

Overall, on average 30% of the cross country gap in TFP appears to be management related (see

base of column (7)).39 This fraction varies a lot between countries. In general we account for a

smaller fraction of the TFP gap between the U.S. and low income countries like Zambia (6.2%),

Ghana (9.7%), and Tanzania (12%), which is likely to be because these countries have much greater

problems than just management quality. We account for a larger fraction of the TFP gap between

the U.S. and richer countries like Sweden (43.9%), Italy (48.9%) and France (52.3%). Figure A4

graphically illustrates this, showing that more developed countries have a higher share of their TFP

gap accounted for by differences in management.

In Appendix B we consider a wide variety of robustness tests of these basic findings, and these

are summarized in Table 8. Row 1 gives the baseline result from Table 7. Row 2 drops pre-2006

data and row 3 drops all panel observations apart from the entry year. We change the selection

equation underlying the sample weights used to correct for non-random response by using only size

(dropping listing status, age and industry dummies) in row 4. Row 5 gives the results without

any selection correction, Row 6 includes multinationals, and Row 7 uses an alternative measure of

size, using an index of weighted inputs (capital and labor). We were concerned that we did not

run our survey on very small (under 50 workers) and very large firms (over 5,000 workers), so we

37Interestingly, these results are broadly consistent with Bartelsman et al (2013) who conducted a similar analysisfor productivity on a smaller number of countries but with larger samples of firms. Although the countries weexamine do not perfectly overlap, the ranking in Bartelsman et al (2013) also has the U.S. at the top with Germanysecond and then France. Britain does somewhat better in our analysis, being above France, but our data is morerecent (2006-2014 compared to their 1992-2001) and Bartelsman et al (2013) note that Britain’s reallocation positionimproved in the 2000s (their footnote 9).

38We used the latest information from 2011, but qualitative results are stable if we take an average over a largernumber of years. When data was missing we impute using values in Jones and Romer (2010).

39For the seven countries where it is possible to calculate manufacturing TFP, the correlation with whole economyTFP is 0.94. The average proportion of the manufacturing TFP gap accounted for by management in these countrieswas 32.6%. We also find that our manufacturing management scores are highly correlated with the managementscores in other sectors like retail, healthcare, schools and government services (see Chong et al. 2014), so that themanufacturing management score appears to be a good measure of overall national management quality.

23

repeated our analysis on a sub-sample of countries where we have detailed information on the firm

size distribution in the population. Knowing the full size distribution allows us to make a selection

correction for the fact we only observe medium sized firms (Appendix B.2). Row 8 has the baseline

results on these countries and row 8 implements the correction.

Although the exact quantitative findings change across Table 8, the qualitative results are very

robust to all these alternative modeling details. The fraction of the TFP gap explained by manage-

ment is non-trivial, ranging from 20% to 50% (column (5)). The correlation of relative management

gaps between the baseline estimates and alternatives estimation techniques (column (3)) never falls

below 0.85 and the correlation of the fraction of TFP explained by management (column (6)) with

baseline results never falls below 0.89.

We can also look at the within country/cross-firm dimension for those countries where we have de-

tailed productivity data. For example, the average industry TFPR spread between the 90th and 10th

percentiles is 90% in U.S. manufacturing (Syverson, 2004), so with our spread of management (2.7

standard deviations between the 90-10) we can account for 30% of the TFP spread (=(2.7*0.1)/0.9).

If instead we examine TFPQ using the results from Foster, Haltiwanger and Syverson (2008) we ob-

tain a similar share of dispersion potentially accounted for by management.40 Similar calculations

for the U.K. show that 38% of the 90-10 TFPR spread is management-related.

6 Conclusions

Economists, business people and many policymakers have long believed that management practices

are an important element in productivity. We collect original cross sectional and panel data on

over 11,000 firms across 34 countries to provide robust firm-level measures of management in

an internationally comparable way. We detail a formal model where our management measures

have “technological” elements. In the model, management enters as another capital stock in the

production function and raises output. We allow entrants to have an idiosyncratic endowment of

managerial ability, but also to endogenously change management over time (alongside other factor

inputs, some of which are also costly to adjust like non-managerial capital). We show how the

qualitative predictions of this model are consistent with the data, as well as presenting structural

estimates to recover some key parameters (such as the cost of adjustment and depreciation rates

of managerial capital).

Our empirical findings are easy to summarize. First, firms who scored more highly in our manage-

ment quality index improved firm performance in both non-experimental and experimental settings.

Second, in the cross section and panel dimension, firms in sectors facing greater competition were

more likely to have better management practices. Part of this competition effect is due to stronger

40While Foster et al. (2008) do not provide data on the 90-10 spread for TFPQ in their data they do provide thestandard deviation which is 0.67 (compared to 0.56 for TFPR) which for a normal distribution would imply a 90-10spread of 95%, implying management would again account for about 30% of the dispersion.

24

reallocation effects, whereby the better managed firms are rewarded with more market share in some

countries compared to others. Third, as cohorts of firms age, the average level of management in-

creases and dispersion decreases (due to selection). Fourth, the falls in the price of management

as proxied by increases in the supply the skills (e.g. through universities and business schools) are

associated with higher management scores. Finally, we use the model to show that management

accounts for about 30% of a nation’s TFP deficit with the U.S. across countries.

There are many directions to take this work. It would be useful to examine the determinants of

management practices in greater detail. We have focused on market-based incentives, but informa-

tional frictions and coordination may be equally if not more important. Gibbons and Henderson

(2012), for example, argue that the need to coordinate a multitude of dispersed agents within a

firm is critical.41

We would also like to test another implication of the management as a technology model, which is

that if management is at least partly non-rival, it should exhibit spillovers as firms learn from each

other. Thus, there will be positive effects of management on those neighbors who can learn best

practice. This is analogous to the R&D or peer effects literature, and techniques can be borrowed

from this body of work as a test of the alternative model, which we leave for future work. Finally,

while we focus on the evidence supporting our “Management as a Technology” interpretation, our

contingency results support a design channel too. So it would be good to develop a richer model

encompassing both the design and technology approaches. We hope our work opens up a research

agenda on why there appear to be so many badly managed firms and what factors can help improve

management, and thus increase the aggregate wealth of nations.

References

Adhvaryu, Achyuta, Namrata Kala and Anant Nyshadham, 2016, “Management and shocks to

worker productivity”, Michigan mimeo.

Akcigit, UfU.K., Harun Alp and Michael Peters. 2016. “Lack of selection and limits to delegation:

firm dynamics in developing countries”, Chicago mimeo.

Aghion, Philippe, Nick Bloom, Richard Blundell, Rachel Griffith, and Peter Howitt. 2005. “Com-

petition and Innovation: An inverted U relationship.” Quarterly Journal of Economics, 120(2):

701-728.

Alessandria, George and Horag Choi. 2007. “Do Sunk Costs of Exporting Matter for Net Export

Dynamics?” Quarterly Journal of Economics, 122(1): 289-336.

Autor, David, Frank Levy and Richard Murnane. 2002. “Upstairs, Downstairs: Computers and

Skills on Two Floors of a Large Bank.” Industrial and Labor Relations Review , 55(3): 432-447.

Atalay, Enghin, Ali Hortascu and Chad Syverson (2014), “Vertical integration and input flows”,

American Economic Review, 104(2).

41The coordination role of CEOs is empirically explored in Bandiera, Hansen, Prat and Sadun (2016).

25

Atkin, Dave, A. Chaudhry, S. Chaudry, Amit Khandelwl and Erik Verhoogen, (2016), “Organiza-

tional barriers to technology adoption: evidence from soccer-ball production in Pakistan”, NBER

Working Paper

Baker, George and Ricard Gil. 2013. “Clinical papers in Organizational Economics” in Robert

Gibbons and John Roberts (eds) Handbook of Organizational Economics. Princeton: Princeton

University Press.

Bandiera, Hansen, Andrea Pratt and Raffaella Sadun. 2016. “CEO Behavior and Firm Perfor-

mance”, HBS mimeo.

Bartelsman, Erik, John Haltiwanger, and Stefano Scarpetta. 2013. “Cross Country Differences in

Productivity: The Role of Allocation and Selection.” American Economic Review, 103(1): 305-334.

Bertrand, Marianne and Antoinette Schoar. 2003. “Managing with Style: The Effect of Managers

on Firm Policies.” Quarterly Journal of Economics, 118(4): 1169-1208.

Bloom, Nick. 2009. “The impact of uncertainty shocks”, Econometrica, 77(3), pp. 623-685.

Bloom, Nicholas and John Van Reenen. 2007. “Measuring and Explaining Management Practices

across Firms and Countries.” Quarterly Journal of Economics, 122(4): 1341-1408.

Bloom, Nicholas, Ben Eifert, Abarjit Mahajan, David McKenzie and John Roberts. 2013. “Does

management matter? Evidence from India.” Quarterly Journal of Economics, 128(1): 1-51.

Bloom, Nicholas, Raffaella Sadun and John Van Reenen. 2012. “Americans do IT Better: American

Multinationals and the Productivity Miracle.” American Economic Review, 102(1): 167-201.

Bloom, Nicholas, Mirko Draca and John Van Reenen. 2015. “Trade Induced Technical Change?

The Impact of Chinese Imports on Innovation, IT and Productivity?” Review of Economic Studies,

83(1): 87-117.

Bloom, Nicholas, Erik Brynjolfsson, Lucia Foster, Ron Jarmin, Megha Patnaik, Itay Saporta-Eksten

and John Van Reenen. 2016. “What drives Management Practices?” Stanford mimeo.

Braguinsky, Serguey, Atsushis Ohyama, Tetsuji Okazaki and Chad Syverson. 2015. “Acquisition,

Productivity and Profitability: Evidence from the Japanese Cotton Spinning Industry.” American

Economic Review, 105(7): 2086-2119.

Bruhn, Miriam, Dean Karlan and Antoinette Schoar. 2010. “What Capital is Missing in Developing

Countries.” American Economic Review, 100(2): 629-633.

Bruhn, Miriam, Dean Karlan and Antoinette Schoar. 2016. “The Impact of Consulting Services

on Small and Medium Enterprises: Evidence from a Randomized Trial in Mexico.” Journal of

Politically Economy, forthcoming.

Bresnahan, Timothy, Erik Brynjolfsson, and Lorin Hitt. 2002. “Information Technology, Work-

place Organization and the Demand for Skilled Labor: Firm-level Evidence.” Quarterly Journal of

Economics, 117(1), 339-376.

26

Card, David, Jorg Heining and Patrick Kline. 2013. “Workplace Heterogeneity and the Rise of

West German Wage Inequality.” Quarterly Journal of Economics, 128(3): 967-1015.

Chirinko, Robert. 1993. “Business Fixed Investment Spending: Modeling Strategies, Empirical

Results, and Policy Implications.” Journal of Economic Literature, 31(4): 1875-1911.

Chong, Alberto, Rafael La Porta, Florencio Lopez-de-Silanes and Andrei Shleifer. 2014. Journal

of the European Economic Association, 12(2), 277-299.

Corrado, Carol and Charles Hulten. 2010. “How do you Measure a Technological Revolution?”

American Economic Review, 100(2): 99-104.

Cyert, Richard and James March. 1963. A Behavioral Theory of the Firm. Englewood Cliffs, NJ:

Prentice-Hall.

Feng, Andrew. 2012. “Essays in Productivity and Management.” London School of Economics,

PhD Thesis.

Feenstra, Robert and John Romalis. 2012. “International Prices and Endogenous Quality.” NBER

Working Papers 18314.

Feenstra Robert, Robert Inklaar and Marcel P. Timmer. 2015. “The Next Generation of the Penn

World Tables.” American Economic Review, 105(10), 3150-3182.

Foster, Lucia, John Haltiwanger and Chad Syverson. 2008. “Reallocation, Firm Turnover, and

Efficiency: Selection on Productivity or Profitability?” American Economic Review, 98(1): 394-

425.

Gennaioli, Nicola, Rafael La Porta, Florencio Lopez-de-Silvanes and Andrei Shleifer (2013) “Human

Capital and Regional Development” Quarterly Journal of Economics 128(1) 105-164

Gibbons, Robert and John Roberts. 2013. The Handbook of Organizational Economics. Princeton:

Princeton University Press.

Gibbons, Robert and Rebecca Henderson. 2012. “Relational Contracts and Organizational Capa-

bilities.” Organizational Science, 23(5): 1350-1364.

Giorcelli, Michela. 2016. “The Long Term Effects of Management and Technology Transfer.”

Stanford mimeo.

Greenstone, Michael, Richard Hornbeck, and Enrico Moretti. 2010. “Identifying Agglomeration

Spillovers: Evidence from Winners and Losers of Large Plant Openings.” Journal of Political

Economy, 118(3): 536-598.

Griliches, Zvi. 1998. R&D and Productivity: The Econometric Evidence. Chicago: Chicago

University Press.

Guner, Nezih and Gustavo Ventura. 2014. “Managers and Productivity Differences.” ASU mimeo.

Hall, Robert and Charles Jones. 1999. “Why Do Some Countries Produce so Much More Output

per Worker than Others?” Quarterly Journal of Economics, 114(1): 83-116.

27

Helpman, Elhanan, Marc Melitz, and Stephen Yeaple. 2004. “Export versus FDI with Heteroge-

neous Firms.” American Economic Review, 94(1): 300-316.

Hopenhayn, Hugo. 1992. “Entry, Exit and Firm Dynamics in Long-Run Equilibrium.” Economet-

rica, 60(5): 1127-50.

Hsieh, Chiang-Tai and Pete Klenow. 2009. “Misallocation and Manufacturing TFP in China and

India.” Quarterly Journal of Economics, 124(4): 1403-1448.

Ichniowski, Casey, Kathryn Shaw, and Giovanna Prennushi. 1997. “The Effects of Human Resource

Management Practices on Productivity: A Study of Steel Finishing Lines.” American Economic

Review, 87(3): 291-313.

Jones, Chad and Paul Romer. 2010. “The New Kaldor Facts: Ideas, Institutions, Population, and

Human Capital.” American Economic Journal: Macroeconomics, 2(1): 224-245.

Lazear, Edwards, Kathryn Shaw and Chris Stanton, 2015, “The value of bosses”, Stanford mimeo.

Lemieux, Thomas, Bentley MacLeod and Daniel Parent, 2009. “Performance pay and wage in-

equality”, Quarterly Journal of Economics, 124(1), pp. 1-49.

Lucas, Robert. 1978. “On the Size Distribution of Business Firms.” Bell Journal of Economics,

9(2): 508-523.

Melitz, Marc. 2003. “The Impact of Trade on Intra-Industry Reallocations and Aggregate Produc-

tivity Growth.” Econometrica, 71(6): 1695-1725.

Milgrom, Paul. 1988. “Employment Contracts, Influence Activities, and Efficient Organization

Design.” Journal of Political Economy, 96(1): 42-60.

Olley, Steve and Ariel Pakes. 1996. “The dynamics of productivity in the telecommunications

industry”, Econometrica, 64(6), 1263-1297.

Prescott, Edward and Michael Visscher. 1980. “Organization Capital.” Journal of Political Econ-

omy, 88(3): 446-61.

Ramey, Valerie and Mathew Shapiro. 2001. “Displaced Capital: A Study of Aerospace Plant

Closings.” Journal of Political Economy, 109(5): 958-992.

Restuccia, Diego and Richard Rogerson. 2008. “Policy Distortions and Aggregate Productivity

with Heterogeneous Establishments.” Review of Economic Dynamics, 11(4): 707-720.

Stokey, Nancy, Robert Lucas and Edward Prescott. 1989. Recursive Methods in Economic Dy-

namics. Cambridge: Harvard University Press.

Syverson, Chad. 2011. “What Determines Productivity?” Journal of Economic Literature, 49(2):

326-365.

Woodward Joan. 1958. Management and Technology. Cambridge: Cambridge University Press.

Walker, Francis. 1887. “The Source of Business Profits.” Quarterly Journal of Economics. 1(3):

265-288.

28

31/05/2016

1

Figure 1: Average Management Scores by Country

Note: Unweighted average management scores; # interviews in right column (total = 15,489); all waves pooled (2004-2014)

2.0272.2212.225

2.2542.316

2.3722.397

2.5162.549

2.5782.6082.611

2.6842.6992.7062.7122.720

2.7482.7522.762

2.8262.8392.8512.861

2.8872.899

2.9782.9973.0153.033

3.1423.188

3.2103.230

3.308

1.5 2 2.5 3 3.5Average Management Scores, Manufacturing

MozambiqueEthiopia

GhanaTanzania

ZambiaMyanmar

NicaraguaNigeriaKenya

ColombiaVietnam

IndiaBrazil

ArgentinaTurkeyChina

GreeceSpainChile

Republic of IrelandPortugal

Northern IrelandNew Zealand

SingaporePolandMexico

ItalyAustralia

FranceGreat Britain

CanadaSweden

GermanyJapan

United States

Africa

Asia

Oceania

Europe

Latin America

North America

1564178749404419

1540780473632406525364151137410161611214585763332568

115115117093718511897

14769

150108131109

Interviews

Figure 2: Management Practice Scores Across Firms

Fra

cti

on

of

Fir

ms

Firm level average management scores, 1 (worst practice) to 5 (best practice)

Note: Bars are the histogram of the actual density. 15,489 interviews.

0.5

11.

50

.51

1.5

0.5

11.

50

.51

1.5

0.5

11.

50

.51

1.5

1 2 3 4 5 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5

Total Argentina Australia Brazil Canada Chile

China Colombia Ethiopia France Germany Ghana

Great Britain Greece India Italy Japan Kenya

Mexico Mozambique Myanmar New Zealand Nicaragua Nigeria

Northern Ireland Poland Portugal Republic of Ireland Singapore Spain

Sweden Tanzania Turkey United States Vietnam Zambia

31/05/2016

2

Figure 3: Management Spreads: Data and Simulation

Notes: The left plot is the management distribution for our entire sample of 15,489 firm surveys. The right plot isthe histogram for a simulation of 15,489 simulated firm-years, where management has been logged and scaledonto a 1 to 5 range. Replication file on http://web.stanford.edu/~nbloom/MAT.zip

0.2

.4.6

Den

sity

1 2 3 4 5Management

0.2

.4.6

Den

sity

1 2 3 4 5lmanagement

Figure 4: Management and Competition - Simulations

Elasticity of Competition

Notes: Results from using our estimated MAT model to simulate 5,000 firms per year in the steady state.Plots log(management) in the simulation data normalized onto a 1 to 5 scale, and log(sales). Competitionis index by demand elasticity (ρ=5) in baseline. Dark Blue bar is unweighted mean across firms, Light Redbar is weighted by firm size (employees). Replication file on http://web.stanford.edu/~nbloom/MAT.zip

Man

agem

ent S

core

1.5

22.

53

3.5

44.

5

3 4 5 6 7 8 10 12 15

Management Size-weighted management

31/05/2016

3

Figure 5: Management and Firm Age - Simulations

Notes: Plots ln(management) scores weighted by age. Results from simulating 5,000 firms per year in thesteady state taking the last 10 years of data and defining age based on the number of observed years.Management normalized to zero on the sample. Replication file on http://web.stanford.edu/~nbloom/MAT.zip

-.5-.2

50

.25

.5M

an

ag

em

en

t m

ean

(nor

mal

ized

to z

ero

mea

n ov

er th

e sa

mpl

e

-1-.5

0.5

1

Man

ag

em

en

t sp

read

(sta

ndar

d de

viat

ions

)

1 2 3 4 5 6 7 8 9 10+

Firm Age

Management level(sample normalized to 0)

Management spread

Figure 6: Management as Design

Notes: Results from the Management as Design model to simulate 5,000 firms in the steady state taking the last 10 years of data.Plots log(management) in the simulation data normalized onto a 1 to 5 scale, and log(sales). See text for more details. Replicationfile on http://web.stanford.edu/~nbloom/MAT.zip

Panel B: Management & Competition

1.5

22.

53

3.5

44.

5

3 4 5 6 7 8 10 12 15

management L weighted management

Elasticity of Competition

Man

agem

ent

Panel D: Management & Performance

7.2

7.4

7.6

7.8

8

1 2 3 4 5

Management

Sale

s (in

logs

)

Panel C: Management & Age

-.25

-.15

-.05

.05

.15

.25

Man

agem

ent m

ean

-.75

-.5-.2

50

.25

.5.7

5

Man

agem

ent s

td

1 2 3 4 5 6 7 8 9 10

Age

Management level(sample normalized to 0)

Management spread

Panel A: The distribution of management

0.2

.4.6

1 2 3 4 5

Management

Den

sity

31/05/2016

4

A) Management & Competition: Levels

Figure 7: Management and Competition

B) Management & Competition: Changes

Notes: Competition proxies are 1-Lerner = median firm profits/sales, Imports = imports/apparent consumption, Imports China = imports from China/apparent consumption, all in industry by country cell. In “levels” panels control for linear country & industry average. “Changes” are in deviations from time-specific country by industry dummies.

Figure 8. Management and Age - Data

Age bins1-5 6-10 11-15 16-20 21-25 26-30 31-35

-.02

0.0

2.0

4

.10

.12

.14

.16

35+

Notes: Data from 31,793 plants from the Management and Organizational Practices supplement to the 2010 Annual Survey of Manufacturing, run by the US Census. Mean management in deviation from the sample mean. Management score from 0 (no modern practices adopted) to 1 (all 16 modern management practices adopted).

Management level(sample normalized to 0)

Management spread

Man

ag

em

en

t m

ean

(nor

mal

ized

to z

ero

mea

n ov

er th

e sa

mpl

e

Man

ag

em

en

t sp

read

(sta

ndar

d de

viat

ions

)

31/05/2016

5

-2.5 -2 -1.5 -1 -.5 0

United StatesJapanSwedenGermanyCanadaGreat BritainSingaporePolandFranceMexicoItalyAustraliaSpainVietnamChileTurkeyPortugalColombiaBrazilIrelandChinaNew ZealandIndiaKenyaGreeceArgentinaMyanmarNigeriaTanzaniaNicaraguaEthiopiaZambiaGhanaMozambique

Figure 9: Management and Reallocation by Country

Notes: Share-weighted management score differences relative to the US (in terms of management score standard deviations).Length of bar shows total deficit, composed of the sum of the (i) the unweighted average management scores (black bar) and theOlley-Pakes reallocation effect (red bar). Domestic firms only with management scores corrected for sampling selection bias.

Reallocation gap with the USUnweighted management score gap with the US

34

TABLE 1: CALIBRATED PARAMETERS FROM THE LITERATURE

Parameter Symbol value Rationale

Capital – output elasticity a 0.3 NIPA factor share Labor – output elasticity b 0.6 NIPA factor share Management – output elasticity c 0.1 Bloom et al (2013) Demand elasticity ρ 5 Bartelsman et al (2013) AR(1) parameter on ln(TFP) ρA 0.885 Cooper and Haltiwanger(2006) Standard deviation of ln(TFP) σA 0.453 Cooper and Haltiwanger(2006)* Discount Factor β 1/1.1 Standard 10% interest rate Capital depreciation rate δK 10% Standard accounting assumption Capital resale loss ϕK 50% Ramey and Shapiro (2001)

Notes: Fixed cost of production is normalized to 100 and mean of ln(TFP) is normalized to 1. * Assumes measurement error on measured TFP has the same variance as true TFP as estimated in Bloom et al (2013).

35

TABLE 2: ESTIMATED PARAMETERS USING SIMULATED METHOD OF MOMENTS

PANEL A: PARAMETER ESTIMATES FROM SMM

Parameter Symbol Value

Depreciation rate of management δM 0.133 (0.055) Adjustment cost parameter for management γM 0.207 (0.065) Adjustment cost parameter for capital γK 0.189 (0.042) Sunk cost of entry κ 165.9 (6.78)

PANEL B: MOMENTS USED IN SMM ESTIMATES

Parameter Simulated Value Data Value

Standard deviation of 5 year management growth 0.559 0.564 Standard deviation of 5 year sales growth 0.936 0.948 Standard deviation of 5 year capital growth 0.883 0.875 Annual Exit rate 3.88% 3.89%

Notes: These are the parameters we estimate using the model (see text). Calibrated parameters from Table 1. Estimation by SMM on the management- accounting panel dataset of up to 4907 firms. Standard errors generated by moment block-bootstrap and moment derivatives taken around the estimated parameter values.

36

TABLE 3: PERFORMANCE REGRESSIONS

(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Dependent variable

Ln (Sales)

Ln (Sales)

Ln (Sales)

Ln (Sales)

Ln (Sales)

Ln (Employees)

Profitability (ROCE, %)

5 year Sales growth

Ln(Tobin Q)

Death (%)

Specification OLS OLS OLS OLS Olley-Pakes OLS OLS OLS OLS Probit Management 0.356*** 0.159*** 0.154*** 0.035*** 0.231*** 0.402*** 1.005*** 0.043*** 0.030** -0.090** (z-score) (0.018) (0.016) (0.019) (0.012) (0.031) (0.013) (0.296) (0.012) (0.015) (0.041) Ln(Employees) 0.873*** 0.623*** 0.623*** 0.427*** 0.524***

(0.019) (0.024) (0.028) (0.060) (0.025) Ln(Capital) 0.306*** 0.295*** 0.186*** 0.386***

(0.019) (0.022) (0.042) (0.098) General controls No Yes Yes Yes Yes Yes Yes Yes Yes Yes

Firm FE No No No Yes No No No No No No Observations 12,146 10,900 8,877 8,877 6,433 24,501 12,578 11,291 6,572 7,507

Notes: *** denotes significance at the 1% level, ** denotes 5% significance and * denotes 10% significance. All columns estimated by OLS (except columns (5) and (10)) with standard errors in parentheses under coefficients and clustered by firm. Column (5) uses the Olley-Pakes estimator. Column (10) uses probit estimation. Columns (1) to (9) are run on the combined full management-accounting dataset panel from 2004 to 2014. Column (10) is run on the 2011 cross-section using firms surveyed up until 2010, with this cut-off defined to provide 4 years to measure exit (firms can take up to 4 years to show up as dead in accounting data). Columns (3) and (4) restrict to firms which were surveyed in at least two different years. Column (5) restricts to firms with lagged capital to generate investment data (the proxy variable in Olley-Pakes). “Management” is the firm’s normalized z-score of management (the z-score of the average z-scores of the 18 management questions). “Profitability” is “Return on Capital Employed” (ROCE) and “5 year Sales growth” is the 5-year growth of sales defined as the difference of current and 5-year lagged logged sales. All columns include a full set of country, three digit industry and time dummies. “Death” is the probability of exit by 2011 (sample mean is 2.4%). “Tobin’s Q” is the stock-market equity and book value of debt value of the firm normalized by the book value of the firm, available for the publicly listed firms only. “General controls” comprises of firm age and the proportion of employees with college degrees (from the survey), plus a set of survey noise controls which are interviewer dummies, the seniority and tenure of the manager who responded, the duration of the interviews and an indicator of the reliability of the information as coded by the interviewer.

37

TABLE 4: COMPETITION AND MANAGEMENT (1) (2) (3) (4) (5) (6) Dependent Variable: Management (z-scored) (1-Lerner) 0.990*** 1.751*** (0.366) (0.443) Import Penetration 0.398** 0.830** (0.170) (0.327) Import Penetration – from China only 2.090** 2.204* (0.972) (1.137) Observations 8,630 8,630 8,630 8,630 8,630 8,630 Size- weight the regressions? No Yes No Yes No Yes Notes: ** indicates significance at 5% level and * at the 10%. The dependent variable in all columns is the z-score of the average z-scores of the 18 management questions. OLS estimates with standard errors in parentheses below coefficients. Standard errors are clustered at the industry by country by period cell in all columns. “Lerner” is the median gross price-cost margin across all firms (from the ORBIS population) in the firm’s three digit SIC industry by country by period cell; “Import penetration” is the value of all imports divided by apparent consumption (total production + imports - exports) in the plant’s ISIC Rev3 industry by country by period cell. “Import penetration – from China only” is the value of all imports originating from China divided by apparent consumption in the plant’s ISIC Rev3 industry by country by period cell. All columns include only firms located in OECD countries and include all data collected across 2004-2014 survey waves. Industry is defined at the 3 digit SIC level in columns (1)-(2), and at the ISIC Rev3 level in columns (3)-(6). All columns include ln(firm employment) and ln(plant employment), plus a full set of country by industry dummies and general controls (firm age and the proportion of employees with college degrees (from the survey), a set of survey noise controls which are interviewer dummies, the seniority and tenure of the manager who responded, the duration of the interviews and an indicator of the reliability of the information as coded by the interviewer). Columns (1) and (2) include a variable measuring the log of the number of firms used to build the Lerner index in the country-industry-period cell. Columns (3)-(6) include the variable measuring apparent consumption in the country-industry-period cell (the denominator of the import penetration variable) in levels. Employment weights are defined with respect to total employment in the industry (3 digit SIC) by country cell.

38

TABLE 5: REALLOCATION TOWARDS BETTER MANAGED FIRMS

(1) (2) (3) (4) (5) (6) Dep. Variable: Employees Employees Employees Employees Employees Employees

Management (M) 182.6*** 268.4*** 294.83*** 285.09*** 463.23*** 289.397*** (US is the omitted base) (20.8) (40.1) (77.22) (45.53) (105.09) (71.538) Management*Employment -56.73* -59.21* Protection (country level) (30.46) (30.66) Management *Trade costs -0.10*** -0.16*** (country level) (0.03) (0.05) Management *Tariff Levels -45.141* (country*industry) (24.653) MNG*Africa -144.6***

(52.1) MNG*Americas -96.3**

(43.9) MNG*(Other EU) -46.6

(58.5) MNG*(South EU) -199.5***

(46.1) MNG*Asia -64.3

(52.3)

Observations 8,991 8,991 7,272 8,918 7,272 8,087

Notes: *** indicates significance at the 1%, 5% (**) or 10% (*) level. Management*US is the omitted base in columns (1) and (2). Estimation by OLS with standard errors clustered by firm. All columns include year, country, three SIC digit industry dummies, and general controls (firm age and the proportion of employees with college degrees (from the survey), plus a set of survey noise controls which are interviewer dummies, the seniority and tenure of the manager who responded, the duration of the interviews and an indicator of the reliability of the information as coded by the interviewer). The sample is domestic firms only (i.e. no foreign or domestic multinationals). “M” is z-score of the average z-scores of the 18 management questions. “Sales growth” is one year ahead logarithmic change. “Employment Protection” is the “Difficulty of Hiring” index is from World Bank (from 1 to 100). “Trade Cost” is World Bank measure of the costs to export in the country (in US$). Tariffs are specific to the industry and country (MFN rates) kindly supplied by John Romalis (see Feenstra and Romalis, 2012).

39

TABLE 6: MANAGEMENT BY DESIGN - STYLES DIFFER DEPENDING ON ENVIRONMENT

(1) (2) (3) (4) (5) (6) (7) (8) (9) People

Management Monitoring & Targets

Relative People

People Management

Monitoring &Targets

Relative People

People Management

Monitoring &Targets

Relative People

(P) (MT) (P-MT) (P) (MT) (P-MT) (P) (MT) (P-MT) Panel A: Using US Four digit industry (NBER, NSF)

ln(K/L) 0.018 0.107*** -0.118*** -0.000 0.096*** -0.125*** (0.015) (0.016) (0.018) (0.014) (0.016) (0.019)

R&D Intensity 0.136** 0.041 0.114 0.031 -0.125* 0.201*** (0.064) (0.087) (0.089) (0.062) (0.072) (0.074)

ln(%degree) 0.139*** 0.123*** 0.011 (0.008) (0.007) (0.010)

Observations 13,681 13,681 13,681 13,681 13,681 13,681 13,681 13,681 13,681

Panel B: Two-Digit industry by county specific value (KLEMS, OECD)

ln(K/L) -0.044 0.039 -0.104** -0.062 0.038 -0.126*** (0.040) (0.032) (0.040) (0.040) (0.034) (0.037)

R&D Intensity 0.496 0.100 0.476* 0.584 -0.004 0.721** (0.358) (0.260) (0.249) (0.413) (0.254) (0.306)

ln(%degree) 0.132*** 0.071*** 0.070*** (0.015) (0.012) (0.019)

Observations 4,855 4,855 4,855 4,855 4,855 4,855 4,855 4,855 4,855 Notes: *** denotes 1% significant, ** denotes 5% significance and * denotes 10% significance. All dependent variables are z-scores of average z-scores of the underlying questions. “People management” is the index built using all management questions between 13 to 18 and “Monitoring and targets” are all the remaining questions. All columns estimated by OLS with standard errors in parentheses under coefficients. All columns control for two-digit SIC industry dummies, country and year dummies, ln(firm employment) and ln(plant employment), plus a full set of general controls (firm age and the proportion of employees with college degrees (from the survey), plus a set of survey noise controls which are interviewer dummies, the seniority and tenure of the manager who responded, the duration of the interviews and an indicator of the reliability of the information as coded by the interviewer). In Panel A the capital-labor ratio is taken from the NBER Bartelsman-Grey dataset and R&D intensity is business R&D divided by employment from NSF. Both capital-labor and R&D intensity are at the four digit level for the US and used across all countries (so no country-specific variation). In Panel B the capital-labor ratio is measured at the two digit by country level from the EU KLEMs dataset and R&D/Value added is from the OECD STAN/ANBERD. EU-KLEMS is only available for a restricted set of countries (Australia, Germany, Italy, Japan, Sweden, UK and US) hence the smaller sample size. Standard errors are clustered at the four digit level in Panel A and two-digit industry by country level in Panel B.

40

TABLE 7: DECOMPOSITION OF CROSS COUNTRY TFP AND MANAGEMENT

(1) (2) (3) (4) (5) (6) (7)

country

Weighted Manage

ment

Un-weighted Manage

ment

Covar- iance

Mng. Gap vs.

US

% Reallo- cation

TFP Gap with US

% TFP due

to Man-agement

US 0.90 0.40 0.50 0 0 Japan 0.57 0.26 0.31 -0.33 56.64 -0.34 9.71 Sweden 0.55 0.38 0.17 -0.35 93.39 -0.08 43.49 Germany 0.36 0.18 0.19 -0.54 57.91 -0.19 28.72 Canada 0.27 0.04 0.24 -0.63 41.92 -0.13 48.64 UK 0.10 -0.11 0.21 -0.81 35.88 -0.15 55.34 Singapore 0.07 -0.23 0.29 -0.84 24.69 Poland 0.04 -0.20 0.23 -0.86 30.69 -0.22 39.26 France -0.01 -0.19 0.18 -0.91 35.29 -0.17 52.52 Mexico -0.07 -0.30 0.23 -0.97 28.21 -0.32 30.20 Australia -0.08 -0.18 0.10 -0.98 40.24 -0.19 51.56 Italy -0.08 -0.18 0.10 -0.98 41.05 -0.20 48.90 Spain -0.14 -0.50 0.36 -1.04 13.19 -0.27 39.03 Vietnam -0.18 -0.50 0.32 -1.08 16.42 Chile -0.19 -0.48 0.29 -1.09 18.81 -0.37 29.48 Turkey -0.20 -0.29 0.09 -1.10 36.92 Portugal -0.22 -0.43 0.20 -1.13 26.27 -0.41 27.41 Colombia -0.23 -0.51 0.28 -1.13 19.28 -0.66 17.21 Brazil -0.26 -0.51 0.25 -1.16 21.20 -0.79 14.63 Ireland -0.29 -0.55 0.26 -1.19 20.33 China -0.31 -0.41 0.11 -1.21 32.49 -0.90 13.44 NZ -0.33 -0.48 0.15 -1.23 28.44 -0.24 51.00 India -0.34 -0.52 0.17 -1.25 26.07 -0.73 17.20 Greece -0.40 -0.52 0.12 -1.30 29.09 -0.35 37.48 Kenya -0.40 -0.55 0.14 -1.30 27.26 -1.39 9.32 Argentina -0.41 -0.55 0.14 -1.31 27.74 -0.38 34.84 Myanmar -0.45 -0.78 0.33 -1.35 12.58 Nigeria -0.61 -0.67 0.06 -1.51 28.85 Tanzania -0.67 -1.04 0.37 -1.57 7.96 -1.34 11.74 Nicaragua -0.71 -0.89 0.19 -1.61 19.44 Ethiopia -0.85 -1.05 0.20 -1.75 17.09 Zambia -0.99 -1.05 0.06 -1.89 23.00 -3.04 6.22 Ghana -1.03 -1.04 0.01 -1.94 25.49 -2.00 9.69 Mzmbique -1.46 -1.60 0.13 -2.36 15.47 -1.10 21.40 Average -1.14 29.68 29.94

Notes: Colum (1) is employment share weighted management z-score. Column (2) is the unweighted score and Column (3) is the management-employment covariance. Column (4) is the weighted management score in column (1) relative to US . Column (5) is the proportion of relative weighted management score in column (1) due to reallocation. The country’s TFP relative to the US in column (6) is from Penn World Tables version 8.1. (Feenstra et al, 2015) available at http://www.rug.nl/research/ggdc/data/pwt/v81/the_next_generation_of_the_penn_world_table.pdf. Column (6) is the fraction of the TFP gap with the US in column (6) that is due to the weighted relative management score in column (1). We assume a one standard deviation increase in the management score causes a 10% increase in TFP (using Table 3 and Bloom et al, 2013). All scores are adjusted for nonrandom selection into the management survey (see text). Only domestic firms used in these calculations (i.e. multinationals and their subsidiaries are dropped).

41

TABLE 8: ROBUSTNESS OF DECOMPOSITIONS (1) (2) (3) (4) (5) (6)

Relative to US size-weighted Management Score Share of TFP gap with US accounted for by size-weighted management

Observations Mean Correlation with baseline Observations Mean Correlation

with baseline 1. Baseline 33 -1.14 1 25 29.90 1 2. Drop pre-2006 data 33 -1.22 0.998 25 33.10 0.992 3. Drop panel firms after

first year 33 -1.17 0.995 25 31.55 0.991

4. Just size in selectionequation 33 -0.99 0.996 25 24.80 0.985

5. No Selection Correction 33 -0.86 0.984 25 19.97 0.873 6. Include multi-nationals

33 -1.02 0.854 25 24.43 0.888 7. Weighted inputs as size

measure 33 -0.98 0.990 25 34.48 0.960

8. Include firms outside50,5000 (baseline) 13 -0.86 1

12 41.16 1

9. Include firms outside50,5000 (corrected) 13 -1.08 0.925

12 49.98 0.968

Notes: This Table shows key statistics for alternative ways of dealing with potential sample selection concerns. The baseline in row 1 summarizes results in Table 7 where we correct for country-specific sample selection (modelled from a first stage probit of whether an eligible firm responded as a function of ln(employment size), whether the firm was publicly listed, and industry dummies). We then weight observations by the inverse of the probability of being in the sample. Row 2 drops all data before 2006. Row 3 drops all observations in the panel after the first year the firm is drawn in the sample. Row 4 uses only ln(employment) in the probit for sample selection. Row 5 uses the raw data, i.e. does not correct for selection. Row 6 includes multinationals (the baseline results only use domestic firms) and includes a foreign multinational dummy in the selection equation. Row 7 uses weighted inputs (capital and labor) as a size measure rather than just labor. Our sampling frame uses firms with between 50 and 5,000 employees. For a sub-sample of countries we know the fraction of employment this covers differs across countries (see Appendix B for details) and can use this to impute management in the “missing” part of the size distribution. We show the results of this in rows 8 and 9. Since this correction can only be done for the sub-sample of countries where we know the full firm size distribution for manufacturing, row 8 shows what the baseline results look like for this sub-sample of countries and row 9 presents the corrected results. “Correlation with baseline” is the correlation of the relevant (country level) variable with the baseline values in row 1. The correlation with size weighted management is in column (3) and the correlation of the fraction of country TFP accounted for by management is in column (6).

A APPENDIX: DATA

We overview the datasets in this Appendix. More information on an earlier version of the dataset

can be found in Bloom, Sadun and Van Reenen (2012b) and Bloom and Van Reenen (2007). More

information on the management survey in general (including datasets, methods and an online

benchmarking tool) is available on http://worldmanagementsurvey.org/.

A.1 Firm-level Accounting Databases

Our manufacturing firm sampling frame

We focus on medium sized manufacturing firms in this paper, so to conduct our surveys we would

ideally draw a sampling frame from a business registry of firms. Unfortunately, the names of these

firms in administrative data from business registers (where it exists) are usually confidential. Three

exceptions were Chile, Colombia and Singapore. In Chile we worked directly with the government,

so it was possible to use the confidential business register data to phone firms - the Industrial

Annual Survey Sample of Firms (Encuesta Nacional Industrial Annual - ENIA), which covers all

manufacturing firms with more than 10 employees. We used a similar model in Colombia, working

with the Business Ministry and World Bank. In Singapore the Ministry of Trade and Industry ran

the survey (with our training) using their population business register.42

For other countries we use publicly available data sources.43 Our main sampling frame for all other

countries was based on various accounting databases supplied by Bureau van Dijk (BVD), a private

sector organization that seeks to compile accounting information on companies from all over the

world. BVD ORBIS has names, industry, addresses, status (e.g. active or bankrupt), and account-

ing information such as employment and sales. ORBIS is constructed from a range of sources,

primarily the national registries of companies (such as Companies House in the U.K.). The full

sources for the sampling frame are listed in column (3) of Table A2.44 In some countries we were

concerned about incomplete coverage by ORBIS so we supplemented it with other sources. We also

used Dun & Bradstreet for Australia and New Zealand. In India we use CMIE Firstsource 2005.45

42Unlike in Chile and Colombia, we were not able to obtain the characteristics of the non-responders in the samplingframe from the Singaporean government due to confidentiality reasons.

43There is usually a fee to be paid for access to the data compiled in an easy to use form, but the names andaddresses of the firms are given in this data (unlike administrative government data where names are not revealed)as they are in the public domain.

44The firm-level databases underlying ORBIS are sometimes packaged under different names in different regionsof the world by BVD (for a given country, the same firms are covered but sometimes with more fields of data). Forexample, in Europe the firm database is called AMADEUS (France, Germany, Greece, Great Britain, Italy, Ireland,Northern Ireland, Poland, Portugal, Spain, Sweden and Turkey); in North America it is called ICARUS (U.S. andCanada) and in parts of Asia it is called ORIANA (China, Japan). Other countries where we simply used ORBISinclude Argentina, Brazil, Mexico and Viet Nam. In Table A2 we simply refer to all of these as ORBIS as this is thesame overall list of firms.

45CMIE is constructed from the Registry of Companies in India. Icarus is constructed from the Dun & Bradstreetdatabase, which is a private database of over 5 million US trading locations built up from credit records, businesstelephone directories and direct research. Oriana is constructed from Huaxia credit in China and Teikoku Database

29

Low income countries pose a particular challenge, as there is generally no well maintained business

register to draw upon. We used a larger variety of data sources to construct sampling frames for

these countries complementing ORBIS with country specific “enterprise maps” put together by

researchers such as John Sutton at the London School of Economics’ International Growth Cen-

tre (IGC) who worked with local consultants. For full details on the African sampling frame, see

Lemos and Scur (2015). Briefly, in addition to ORBIS and the enterprise maps in Ethiopia (Sut-

ton and Kellow, 2010), we also used trade directories and the online yellow pages; in Ghana we

supplemented Sutton and Kpentey (2012) with the Ghana Free Zones Board and Business Ghana

Directory; in Kenya we supplemented ORBIS with the Kenya Association of Manufacturers; in

Mozambique we supplemented Sutton, Pimpao, Simione, Zhang and Zita (2014) with the Ministry

of Commerce’s database and FACIM; in Nigeria we supplemented ORBIS with VConnect; in Tan-

zania we supplemented Sutton and Olomi (2012) with the Ministry of Trade database; in Zambia

we supplemented Sutton and Langmead (2013) with the Zambia Association of Manufacturers. In

Myanmar we supplemented the PEDL Enterprise Map with industry directories and lists of manu-

facturers provided by the Myanmar Industry Association (see Tanaka, 2015) and in Nicaragua we

used business directories and an IADB database.

These databases all provide sufficient information on companies to conduct a stratified telephone

survey (company name, address, and a size indicator). They also typically have some accounting

information on employment, sales, and capital. We did not insist on having accounting information

to form the sampling population, however.

Representativeness of the sampling frame

How representative are our sampling frames of the underlying population of manufacturing firms?

For most of the countries we have evidence that the data is reasonably comprehensive. For example,

when comparing aggregate employment in the ORBIS populations to those from census data, we

usually find a reasonably close match (e.g. Kalemli-Ozcan et al, 2015; Bloom, Draca and Van

Reenen, 2015).

In Bloom, Sadun and Van Reenen (2012b) we analyze this in more detail. For example, we compare

the number of employees for different size bands from our sampling frame with the figures for the

corresponding manufacturing populations obtained from national census data from each of the

countries where this is available. There are several reasons for a mismatch between Census data

and firm level accounts.46 Despite these potential differences, our sampling frame appears to cover

in Japan, covering all public and all private firms with one of the following: 150 or more employees, 10 million US$of sales, or 20 million US$ of assets.

46First, even though we only use unconsolidated firm accounts, employment may include some jobs in overseasbranches. Second, the time of when employment is recorded in a census year will differ from that recorded infirm accounts. Third, the precise definition of “enterprise” in the census may not correspond to the “firm” incompany accounts. Fourth, we keep firms whose primary industry is manufacturing whereas census data includes onlyplants whose primary industry code is manufacturing. Fifth, there may be duplication of employment in accountingdatabases due to the treatment of consolidated accounts. Finally, reporting of employment is not mandatory for the

30

near to the population of all firms for most countries. 47

Mismatch between the true population and our sampling frame could be a problem if our sampling

frame was non-randomly omitting firms - for example under-representing smaller firms - because it

would bias our cross-country comparisons. We tried several approaches to address this. First, in

almost all the regression tables we include country fixed effects to control for any differences across

countries in sample selection bias. Hence, our key results are identified by within country variation.

Second, we ran experiments where we dropped problematic countries (e.g. Portugal and Sweden)

from the analysis to show that the results are robust.

It is harder to make such comparisons between our sampling frame and the full population of firms

in some of the poorer countries as there is no reliable employment aggregate numbers to compare

to. The enterprise maps by Sutton and others that we draw our sampling frames from are probably

the most reliable sources. Given this, we should interpret some of the results from the African

countries with more caution than those of the other nations.

Medium sized manufacturing firms vs. all manufacturing firms

A further concern that is that even if our sampling frame is fully comprehensive, the proportion of

employment covered by medium sized firms differs systematically across countries. Using Census

sources on firm populations, Table A3 shows the employment distribution for the countries where

it is available. Firms employing between 50 and 5,000 workers account for about half of all man-

ufacturing workers in most countries, although the proportion was larger in some countries such

as Ireland (72%) and Poland (71%). The proportion employed by very large firms with over 5,000

workers varies more between nations. The patterns are broadly consistent with our MAT model.

In countries where competition is strong and reallocation easier, there is a larger fraction of jobs

in very large firms (e.g. 34.7% in the U.S.) and a small fraction in small firms with under 50

employees (e.g. 16.2% in the U.S.). Germany, also a high productivity and high management score

country, looks similar to the U.S. (34.9% in large firms vs. 16.5% in small firms). By contrast, in

countries like Italy and Greece, only 6.4% and 6.2% of employees respectively are in these large

firms compared to 45.1% and 41.3% in small firms. In Appendix B we provide some corrections to

our cross country decompositions to deal with the fact that we are missing management scores in

the very large firms in some countries (e.g. the U.S. and Germany) compared to other countries,

like those in southern Europe. As expected, these corrections strengthen our overall findings as

the “corrected” relative average management scores for the U.S. and Germany rise compared to

southern Europe.

accounts of all firms in all countries. This was particularly a problem for Indian and Japanese firms, so for thesecountries we imputed the missing employment numbers based on a regression of sales on employment for firms wherewe had both variables.

47In two countries the coverage from accounting databases underestimates the aggregate: the Swedish data coversonly 62% of Census data and the Portuguese accounting database covers 72%. This is due to incomplete coverage inORBIS of these smaller nations.

31

A caveat to Table A3 is that total employment in firms with over 5,000 workers is not disclosed

in all countries (because of concerns it would reveal individual firm’s identities). In the U.S. and

Japan we have the exact Census numbers from public use tables and in the U.K. we had access

to confidential Census micro-data to do this ourselves. In the other countries we used accounting

data from ORBIS and other sources to estimate employment for the very large firms. Since these

firms are so large, data is relatively plentiful as they are almost all publicly listed and so followed

closely by market analysts. 48

A.2 The World Management Survey

In every country the sampling frame for the management survey was all firms with a manufacturing

primary industry code and that employed between 50 and 5,000 workers49 on average over the most

recent three years of data prior to the survey.50Interviewers were each given a randomly selected list

of firms from the sampling frame. The size of this sampling frame by country is shown in column

(5) of Table A2 for the first year that we interviewed a firm. We have conducted the surveys over

multiple years as noted in column (6) of Table A2. The five major waves were in 2004, 2006,

2009/10, 2013 and 201451, although we had smaller scale surveys in some of the intervening years

(e.g. China in 2007; Brazil, Canada and Ireland52 in 2008).

In the first survey in 2004 we covered 732 firms in France, Germany, the U.K. and the U.S.53 In

2006 we covered eight countries (China, Greece, India, Italy, Japan, Poland, Portugal and Sweden)

as well as the four core countries. In addition to the new countries and a refreshment sample

of the four 2004 countries we also re-contacted all firms from 2004 to form a short panel. In

2009/10 we again resurveyed all firms interviewed 2006 including the original 2004 firms (if they

were still alive). For budgetary reasons we did not do a refreshment sample in this wave although

48Corrections have to be made to estimate the number of domestic employees (which is the Census concept) if thisis not revealed directly by the firms. To do this we ran country specific regressions of the proportion of domestic overtotal global employment on a polynomial of total employment, industry dummies, and multinational status. Thenwe used this to impute the number of domestic workers for the firms who did not disclose domestic employment.

49In Japan and China we used all manufacturing firms with 150 to 5000 employees since Oriana only samplesfirms with over 150 employees.Note that the Oriana database does include firms with less than 150 employees if theymeet the sales or assets criteria, but we excluded this to avoid using a selected sample. We checked the results byconditioning on common size bands (above 150 in all countries) to ensure that the results were robust.

50In the U.S. only the most recent year of employment is provided. In India, employment is not reported for privatefirms, so for these companies we used forecast employment, predicted from their total assets (which are reported)using the coefficients from regressing ln(employees) on ln(assets) for public firms.

51Major waves were started in early summer, but sometimes stretched throughout the year. In 2009, the wavestretched through to the following February. We kept information on when the interview took place to control forany seasonal influences (a noise control).

52We split out Northern Ireland from the rest of the U.K. in Table A2 as we did an additional wave specifically ofNorthern Irish firms in 2008. Some of the Northern Irish firms were also surveyed in 2004, 2006 and 2010 as part ofthe general U.K. waves, but only a smaller number as the region is only a small part of the U.K..

53This sample was drawn from the BVD Amadeus dataset for Europe and the Compustat dataset for the U.S. Onlycompanies with accounting data were selected. So, for the U.K. and France, this sampling frame was very similarto the 2006 sampling frame. For Germany it is more heavily skewed towards publicly quoted firms since smallerprivately held firms report little balance sheet information. For the U.S. it comprised only publicly quoted firms. Asa robustness test we drop the firms that were resurveyed from 2004.

32

we did add New Zealand and Australia. In 2013 we mainly surveyed low income countries in Africa

(Ethiopia, Ghana, Mozambique, Nigeria, Tanzania and Zambia) and Latin America (Colombia and

Nicaragua) for the first time. But we also followed Argentina, Mexico and Brazil in the panel and

surveyed Spain and Turkey for the first time. In 2014 we included new countries (Kenya, Myanmar

and Vietnam) and performed refreshment samples of the U.S. and the main EU countries (France.

Germany, Greece, Italy, Portugal and the U.K.). This included attempting to re-survey all the

panel firms from the these EU countries and the U.S. that we had from earlier waves.

The Survey response rate

Column (4) of Table A2 shows the response rates by country. Of the firms we contacted 41% took

part in the survey, a high success rate given the voluntary nature of participation. Of the remaining

firms, 16% refused to be surveyed, while the remaining 43% were in the process of being scheduled

when the survey ended.

There are clear differences by country, with the response rate being lowest in Japan and highest in

Nigeria. Table A4 analyzes the probability of being interviewed.54

All columns condition on country dummies, industry dummies and time dummies (for year of

interview and year the accounting data is taken from if different). The first column simply includes

employment as a measure of size. The size variable is significant suggesting that there was a

tendency for large firms to respond more frequently, although the coefficient is not large: a doubling

of firm size is associated with only a 6% higher probability of response.

In column (2) we add in labor productivity and reassuringly find that firms with higher revenues

per employee are not significantly more likely to agree to be interviewed than others. Column (3)

includes the return on capital employed (ROCE) as an alternative performance measure which is

also insignificant. These are important results as they suggest we are not interviewing particularly

high or low performing firms. In columns (4), (5) and (6) we find that multinational status, firm

age, and capital are also all insignificant in the selection equations.

So, in summary, within a country, industry, and year, respondents were not significantly more

productive, profitable, or capital intensive than non responders. Respondents did tend to be slightly

larger, but were not more likely to be older or multinationals. Since all regressions include size,

country, industry, and time dummies, this potential source of bias is controlled for.

One concern, however, is that when we make cross country comparisons these selection effects

could create biases. Hence in our decomposition analysis we re-weight the data to allow for country

specific sampling bias (see Appendix B and main text). Specifically, we ran probits of whether a

54The responders sample is smaller than the total number of observations in Figure 1 for three reasons. First, welook only at the first time we contact a firm: this is because firms that appear in the panel are likely yo be bettermanaged and more productive (since they survived longer). Second, we conduct the analysis at the levels of the firm’saccounts and there may be multiple plants analyzed for the same firm in the same year. Third, we do not includeSingapore because (as noted above) we do not have access to the non-responders.

33

firm responded compared to others in the sampling frame where the right hand side variables are

employment, firm age, industry dummies and time dummies. We then calculated the probability

of sample response and used the inverse of the sampling probability as weights (winsorizing the top

and bottom percentiles).

Management panel data

As described above, we have a panel element to the database. This is built by combining the

management survey waves with the accounting panel data, and then interpolating (but never ex-

trapolating) the management data to fill in the missing years. For example, if we measure man-

agement practices every three years, we linearly interpolate the data in the intervening years. This

helps to increase the sample sizes for the five year difference moment since we do not need to have

exactly five years between survey waves (for example, without interpolating, if we survey firms

every three years we can never generate a five year difference). But, to confirm our results are

not dependent on interpolating, we can replicate Table 3 using non-interpolated data finding very

similar results.55All regression standard errors are clustered by firm so that the standard-errors are

appropriately corrected for any potential serial correlation in the errors induced by interpolation.

In terms of sample size we have 6,114 firm-year observations between 2004 and 2014 inclusive

for the non-interpolated data and 12,146 firm-year observations between 2004 and 2014 inclusive

for the interpolated data, so that about 49.7% of the firm-years have interpolated management

data (the accounts data is never interpolated). Finally, we also have about 5% of firm-years with

multiple survey values, for which we take the average value of all continuous variables (including all

the overall management score) and the value from the chronologically first survey for the discrete

values (e.g. “plant county of location”, “ownership type” etc). The panel covers 15 countries -

Brazil, Canada, China, France, Germany, Greece, India, Ireland, Italy, Japan, Poland, Portugal,

Sweden, UK and US.

Firm-level variables

We have firm accounting data on sales, employment, capital, profits, shareholder equity, long-term

debt, market values (for quoted firms), and wages. We also collected information on whether

the firm was part of a multinational enterprise during the survey interview. We supplemented

and verified this information through BVD, web searches, and direct telephone inquiries to the

firms. We also collected specific questions on the multinational status of the firm (whether it

owned plants abroad and the country where the parent company is headquartered) to be able to

distinguish domestic multinationals from foreign multinationals.

55For example, the coefficients (standard errors) on management for columns (1) to (4) from Table 3 with non-interpolated data are 0.274 (0.017), 0.142 (0.015), 0.123 (0.016) and 0.038 (0.014) respectively.

34

We collected many variables through our survey, including information on plant size, skills, orga-

nizational structure, etc. as described in the main text. We asked managers to estimate how many

competitors they thought they faced, and used this information to construct one of the firm level

competition variables.

Management practices were scored following the methodology of Bloom and Van Reenen (2007),

with practices grouped into four areas: operations (three practices), monitoring (five practices),

targets (five practices), and incentives (five practices). The shop-floor operations section focuses on

the introduction of lean manufacturing techniques, the documentation of processes improvements,

and the rationale behind introductions of improvements. The monitoring section focuses on the

tracking of performance of individuals, reviewing performance, and consequence management.56

The targets section examines the type of targets, the realism of the targets, the transparency

of targets, and the range and interconnection of targets. Finally, the incentives section includes

promotion criteria, pay and bonuses, and fixing or firing bad performers, where best practice

is deemed the approach that gives strong rewards for those with both ability and effort. Our

management measure averages the z-scores of all 18 dimensions and then z-scores this average

again.

Industry level variables

Our basic industry code is the U.S. SIC (1997) three digit level - which is our common industry

definition in all countries except otherwise noted. We allocate each firm to its main three digit

sector (based on sales), covering 135 unique three-digit industries. There are at least ten sampled

firms in each industry for 97% of the sample.

The “Lerner index of competition” is constructed, as in Aghion et al (2005), as the cross-firm

median of (1-profit/sales) in a country-industry-time period cell. Profits are defined as EBIT

(earnings before interest and taxation) to include the costs of labor, materials, and capital, but

exclude any financing or tax costs. The source of accounting data used to build the index is ORBIS

population data for all countries (i.e. we do not condition on having any management data). We

first build the median Lerner index for every country-industry-year cell.57 We then average the

yearly values over three three-years sub-periods: period 1, including years between 2004 and 2006;

period 2, including years between 2008 and 2010; and period 3, including 2012 and 2013 data (2013

being the last year for which we have full ORBIS extractions). This timing roughly corresponds

to pre-crisis, crisis and post-crisis periods. These time varying measures of the Lerner index are

matched to the management interviews according to the year in which these were conducted (2004-

2006 for period 1; 2008-2010 for period 2; and 2012-2014 for period 3).

The variable “Import penetration” is built using data drawn from the World Input-Output Database

56Since the operations and monitoring concepts overlap we often group them together as “monitoring”.57To reduce the influence of outliers, we drop observations below the 1th and above the 99th percentile of the

Lerner index.

35

(WIOD). Import penetration is built as the share of imports over apparent consumption (total

production-exports+imports). This measure is built at the country by industry by year level and

then, similarly to the Lerner index, averaged across sub-periods. Since the last available year for the

WIOD is 2011, we use only two sub-periods, 2004-2006 (matched with management data collected

between 2004-2006) and 2008-2010 (matched with management data collected between 2008-2014).

The industry classification is ISIC, Rev 3.

The variable “Import penetration from China” is built exactly as the “Import Penetration” variable,

except that it includes only imports originating from China in the numerator.

Tariff data was kindly supplied by Feenstra and Romalis (2015) and is at the country by industry

(SITC, Rev 4) by year (2000-2006) level. Since we have limited time series variation, we collapse

the data to the country by industry level.

Country level variables

We take country level measures of GDP per capita and TFP from the latest Penn World Tables

(Timmer et al, 2015). The OECD “Difficulty of Hiring” index (ranging from 1 to 100) is the variable

EPRC V2 (available for all years in the interval 1998-2013), and is drawn from their indicators of

employment protection. These are synthetic indicators of the strictness of regulation on dismissals

and the use of temporary contracts, compiled from 21 items covering three different aspects of

employment protection regulations as they were in force on January 1st of each year. We also

employ the variable “Cost to export (US$ per container)” drawn from the World Bank “Doing

Business” dataset (2015) which reflects the cost of obtaining export licenses and average container

costs.

A.3 Management and Organizational Practices Survey (MOPS)

MOPS was jointly funded by the U.S. Census Bureau and the National Science Foundation as

a supplement to the Annual Survey of Manufactures (ASM). The original design was based on

the same concepts as the WMS and was adapted to the U.S. through several months of devel-

opment and cognitive testing by the Census Bureau. It was sent by mail and electronically to

the ASM respondent for each establishment, which was typically the accounting, establishment, or

human-resource manager. Most respondents (58.4%) completed the survey electronically, with the

remainder completing the survey by paper (41.6%). Non-respondents were given up to three follow-

up telephone calls if no response had been received within three months. The survey comprised 36

multiple choice questions about the establishment, taking about 20 to 30 minutes to complete. The

survey included 16 questions on management practices covering (like WMS) monitoring, targets

and incentives.

The monitoring section asked firms about their collection and use of information to monitor and

improve the production process. For example, firms were asked how frequently performance indi-

cators were tracked at the establishment, with options ranging from “never” to “hourly or more

36

frequently”. The targets section asked about the design, integration, and realism of production

targets. For example, firms were asked what the time-frame of production targets was, ranging

from “no production targets” to “combination of short-term and long-term production targets”.

Finally, the incentives section asked about non-managerial and managerial bonus, promotion and

reassignment/dismissal practices. For example, it asked how managers were promoted at the es-

tablishment, with answers ranging from “mainly on factors other than performance and ability, for

example tenure or family connections” to “solely on performance and ability”. The full question-

naire is available on http://bhs.econ.census.gov/bhs/mops/form.html.

In our analysis, we aggregate the results from these 16 check box questions into a single measure of

structured management. The structured management score is the unweighted average of the score

for each of the 16 questions, where each question is first normalized to be on a 0-1 scale. Thus

the summary measure is scaled from 0 to 1, with 0 representing an establishment that selected the

bottom category (little structure around performance monitoring, targets, and incentives) on all

16 management dimensions, and a 1 representing an establishment that selected the top category

(an explicit focus on performance monitoring, detailed targets, and strong performance incentives)

on all 16 dimensions. As with WMS, we asked a range of questions about the characteristics of the

workers, firms, and collected variables that we used to approximate for interview noise (interviewee,

interview and interviewer characteristics).

The MOPS survey was sent to all ASM establishments in the ASM mail-out sample. 37,177

filled surveys were received, implying a response rate of 78%, which is extremely high for firm

surveys.We further restricted the sample for establishments with at least 11 non-missing responses

to management questions and also have positive value added, positive employment and positive

imputed capital in the ASM. In addition to our management data we use establishment level data

on sales, value-added and labor inputs from the ASM. The mean establishment size is 167 employees

and the median is 80. The average establishment in our sample has been in operation for 22 years,

44% of managers and 9% of non-managers have college degrees, 13% of their workers are in unions,

42% export and 69% are part of larger multi-plant firms.

More details on the MOPs survey are contained in Bloom et al (2016).

37

B APPENDIX: FURTHER RESULTS

In Section 5 we decompose share-weighted management into reallocation and unweighted average

components and use this to estimate the contribution to cross country TFP. In doing that, we

made a variety of assumptions that we now relax to see if they materially alter our results. Recall

that the sample we used in the analysis is a sub-sample of that underlying Figure 1 as we drop

multinationals because it is unclear what the appropriate measure of size is for such firms (we also

show robustness to including multinationals). Also, we weight the management data according to

a firm’s country-specific employment share and adjust for non-random selection through sampling

weights.

B.1 Differential response rates to the survey

There are several potential sources of sample selection, the most obvious one being that the firms

who responded from the sampling frame were non-random in some dimension. Appendix A exam-

ined the overall evidence on sampling bias and argued that these were relatively small both on the

observable and unobservable dimensions. Nevertheless, the baseline results in Table 7 control for

this by calculating (country-specific) weights for the sample response probabilities. We do this by

running country-specific probit models where the control variables are ln(employment size), firm

age, a dummy for whether the firm was publicly listed and industry dummies. We chose these con-

trols because they are available for responders and non-responders, and there was some evidence

from Appendix A that larger firms were more likely to respond. We then calculated the weights as

the inverse of the probability of response.

We experimented with an alternative first stage probit to look into sample response bias, based

on just using employment rather than the larger set of controls. The results are summarized in

Table 8. Row 1 presents a summary of the baseline results that we use in Table 7. Row 2 shows

what happens when we drop pre-2006 data. This is motivated by the fact that 2004 was our first

(and smallest) survey wave. The fraction of the TFP gap accounted for by management rises to

33% (from 30%). We also show in column (6) that the correlation of this fraction explained with

the baseline in row 1 is very high (0.992) and the correlation of the gaps in management with the

baseline of Row 1 are also very high (0.998 in column (3)). Row 3 drops all observations which are

in the panel part of the dataset except the first year. The motivation for doing this is that firms who

have been in the panel for one year are more likely to respond in subsequent years than a randomly

chosen firm. Since we attempt to re-sample panel firms in subsequent years this could potentially

generate a bias. In the baseline results of Row 1 we down weight these firms appropriately using

estimate from the probits. Row 3 shows that our results are robust to just dropping them from

the sample (32% of TFP accounted for). We also looked at the sensitivity of the first stage probit

to the controls by dropping all covariates except size (row 4) and not using any sample weights

(row 5). Not using any selection correction generates the smallest fraction of the TFP explained in

38

Table 8 (20%).

In Table 7, we dropped multinationals because of the difficulty of measuring group size appropriately

for such companies. To check the robustness of these results, we included them in row 6, but also

included multinational status into the selection equation used to calculate the sample response rate

weights (multinationals were slightly more likely to participate in the survey). We account for a

bit less of the average TFP gap (24% compared to 30% in the baseline) and the cross country

correlation remains reasonably high (0.89).

We have focused on employment as our key measure of size as it is simple, a volume indicator which

is straightforward to measure across countries. An alternative way to measure size is to look at a

measure of weighted inputs, so we follow Bartelsman et al (2013) and construct a measure using

capital stock information from Orbis where our composite input measure was exp[0.7*ln(labor) +

0.3*ln(capital)]. The results are in row 7 of Table 8 and are again are similar to the baseline,

although with slightly more of TFP accounted for by management (34.5%).

B.2 Sampling biases associated with dropping very small and very large firms

Our management surveys focus on medium sized firms defined as those with over 50 and under

5000 employees. This was in order to compare firms of a broadly similar size. However, it could

potentially cause bias in our comparisons of management levels across countries as the size distri-

bution is different across nations (e.g. Garicano, Lelarge and Van Reenen, 2013, and Table A3).

Obviously we do not know the exact distribution of management scores in these very large and

very small firms, but we can estimate with additional assumptions what the potential biases could

be.

From the census manufacturing population databases of firm demographics, in most countries we

know the number of firms above and below 50 employees, and the total number of workers employed

across firms of different sizes (see Table A3). We need to then make an assumption about the

relationship between firm size and management for the very large and very small firms, which we

extrapolate off the size-management relationship over the part of the distribution that we observe

(50 to 5,000 employees). We checked that the extrapolated size-management relationship holds for

firms below 50 and above 5,000 using the MOPs dataset which asks management questions to firms

from all parts of the U.S. size distribution (Bloom et al, 2016).58

We then use this information to estimate the weighted average management score across the entire

distribution. Our preferred method exploits the fact that the firm size distribution in each country

follows a power law (Axtell, 2001). Using results from this literature we can approximate the

employment weighted mean management score in the sub-population under 50 workers and the

58The coefficient on ln(employment size) in the management regression is 0.25. We considered imposing a commonconstant on each country (-1.46) or adjusting this to be consistent with the country-mean management score in the50 to 5,000 range. Both methods lead to similar results.

39

sub-population with over 5,000 workers.59 We then use the information in Table A3 to calculate

the mean management score across the entire size distribution. Results of this exercise are in rows

8 and 9 of Table 8. We first reproduce the baseline results on the smaller sample of countries in

row 8 of Table 8. In this sub-sample of 14 countries (13 differences as the U.S. is the base) we

account for a higher fraction of TFP (41%) since these are OECD countries. We then implement

the size correction. The correlation between our baseline management scores and the new “size

corrected” management scores is very high (0.925), as is the correlation of the fraction of TFP

explained (0.968). If anything, we account for an even higher fraction of the TFP gap (just under

50%) when making this correction.

B.3 Manufacturing TFP vs. Whole Economy TFP

The management scores are derived from manufacturing firms, but the decompositions in Table

7 use economy wide TFP. This is purely due to data restrictions - there are few countries in

the world for which it is possible to construct reliable measures of manufacturing TFP that are

comparable across countries. Our assumption is that countries with high economy wide TFP also

have high manufacturing TFP. To test whether this might bias our results, we draw on estimates of

manufacturing specific TFP levels (Citino, Haskel and Van Reenen, 2016). This is based primarily

on the KLEMs data and constructs TFP in the same way as the Penn World Tables, correcting for

capital services and skills. There are only seven countries that overlap between the two exercises

where data is sufficiently rich: Australia, Germany, Italy, Spain, Sweden, the U.K. and U.S. The

correlation of manufacturing TFP (relative to the U.S.) for these countries with economy TFP is

reassuringly high: 0.938. The average proportion of the TFP gap accounted for by management is

35%, similar to the overall mean (but smaller than the proportion for just the seven countries in

the baseline which is 43%).60

C APPENDIX: SMM DETAILS

SMM starts by selecting an arbitrary starting value of the parameter vector to be estimated (θ).

The dynamic program is then solved and the policy functions are generated. These policy functions

are used to create a simulated data panel of size (µN, T ), where µ is a strictly positive integer, N

is the number of firms in the actual data, and T is the time dimension of the actual data. The

59First, we consider the approximation in Johnson et al (1994) showing that the number of employees in each size“bin” is equal when the bins are logarithmically sized if firm size is distributed according to Zipf’s Law (which isapproximately true in the data). We predict management in each bin and then employment weight the bin to obtainmean management for the below 50 and above 5,000 firms. We also considered the continuous version of the powerlaw which lead to similar results.

60We performed a similar exercise for the market sector (i.e. dropping health, education and public administration).For this sample we do not have Australia, but we do can add France. Again the correlation of TFP between thissector (market sector) and total economy is high at 0.844. We could account for 43.6% of the cross country TFPgaps, very similar to the fraction for these seven countries in Table 7 (44.7%).

40

simulated moments ΨS(θ) are calculated on the simulated data panel, along with an associated

criterion function Γ(θ), where Γ(θ) = [ΨA−ΨS(θ)]′W [ΨA−ΨS(θ)], which is a weighted distance61

between the simulated moments ΨS(θ) and the actual moments ΨA.

θ = arg minθ∈Θ

[ΨA −ΨS(θ)]′W [ΨA −ΨS(θ)] (7)

A second parameter value is then drawn by taking a random jump from the first value, and the third

parameter value onwards is drawn by taking a random jump away from the best prior guess (the

parameter value that has delivered the lowest criterion function up to that point). This way, the

parameter estimate θ derived by randomly searching over the parameter space to find the parameter

vector, which minimizes the criterion function. This simulated annealing random jumping approach

is used because of the potential for discontinuities in the model and the discretization of the state

space (so a gradient minimization approach may simply find a local rather than a global minimum).

Finally, different initial values of θ are selected to ensure the solution converges to the global

minimum.

References

[1] Axtell, Robert. 2001. “Zipf Distribution of U.S. Firm Sizes.” Science, 293(5536): 1818-1820.

[2] Bloom, Nicholas, Raffaella Sadun, and John Van Reenen. 2012b. “The Organization of Firms

Across Countries.” Quarterly Journal of Economics 127(4): 1663-1705.

[3] Citino, Luca, Jonathan Haskel and John Van Reenen. 2016. “Accounting for the U.K. Pro-

ductivity Gap.” London School of Economics mimeo.

[4] Garicano, Luis, Claire Lelarge and John Van Reenen. 2013. “Firm Size Distortions and the

Productivity Distribution: Evidence from France.” NBER Working Paper No. 18841.

[5] Johnson, Norman, Adrienne Kemp and Samuel Kotz. 1993. Continuous Univariate Distribu-

tions. New York, N.Y.: John Wiley & Sons.

[6] Kalemli-Ozcan, Sebnem, Bent Sorensen, Carolina Villegas-Sanchez, Vadym Volosovych, Sev-

can Yesiltas (2015) “How to Construct Nationally Representative Firm level data from the

ORBIS Global Database” NBER Working Paper 21558.

[7] Lemos, Renata and Daniela Scur. 2015. “A snapshot of mid-sized firms in Africa, Asia

and Latin America.” Oxford University mimeo, http://worldmanagementsurvey.org/wp-

content/images/2010/08/lemos scur snapshot dec20151.pdf.

61The efficient choice for W is the inverse of the variance-covariance matrix of [ΨA−ΨS(θ)], which Lee and Ingram(1991) show under the null can be calculated from the variance-covariance of the empirical moments.

41

[8] Sutton, John and Bennet Kpentey. 2012. Enterprise Map of Kenya. London: IGC.

[9] Sutton, John, Jeque Pimpao, Felix Simione, Qi Zhang and Samuel Zita. 2014. Enterprise Map

of Ghana. London: IGC.

[10] Sutton, John and Donath Olomi. 2012. Enterprise Map of Mozambique. London: IGC/

[11] Sutton, John and Gillian Langmead. 2013. Enterprise Map of Zambia., London: IGC, London:

IGC.

[12] Tanaka, Mari. 2015. “Exporting Sweatshops: Evidence from Myanmar.” Stanford University

mimeo.

42

42

TABLE A1: MANAGEMENT PRACTICES QUESTIONNAIRE

Any score from 1 to 5 can be given, but the scoring guide and examples are only provided for scores of 1, 3 and 5. The survey also includes a set of Questions that are asked to score each dimension, which are included in Bloom and Van Reenen (2007).

(1) Modern manufacturing, introduction Score 1 Score 3 Score 5

Scoring grid: Other than Just-In-Time (JIT) delivery from suppliers few modern manufacturing techniques have been introduced, (or have been introduced in an ad-hoc manner)

Some aspects of modern manufacturing techniques have been introduced, through informal/isolated change programs

All major aspects of modern manufacturing have been introduced (Just-In-Time, autonomation, flexible manpower, support systems, attitudes and behaviour) in a formal way

(2) Modern manufacturing, rationale Score 1 Score 3 Score 5

Scoring grid: Modern manufacturing techniques were introduced because others were using them.

Modern manufacturing techniques were introduced to reduce costs

Modern manufacturing techniques were introduced to enable us to meet our business objectives (including costs)

(3) Process problem documentation Score 1 Score 3 Score 5

Scoring grid: No, process improvements are made when problems occur.

Improvements are made in one week workshops involving all staff, to improve performance in their area of the plant

Exposing problems in a structured way is integral to individuals’ responsibilities and resolution occurs as a part of normal business processes rather than by extraordinary effort/teams

(4) Performance tracking Score 1 Score 3 Score 5

Scoring grid: Measures tracked do not indicate directly if overall business objectives are being met. Tracking is an ad-hoc process (certain processes aren’t tracked at all)

Most key performance indicators are tracked formally. Tracking is overseen by senior management.

Performance is continuously tracked and communicated, both formally and informally, to all staff using a range of visual management tools.

43

(5) Performance review

Score 1 Score 3 Score 5 Scoring grid: Performance is reviewed infrequently or in

an un-meaningful way, e.g. only success or failure is noted.

Performance is reviewed periodically with successes and failures identified. Results are communicated to senior management. No clear follow-up plan is adopted.

Performance is continually reviewed, based on indicators tracked. All aspects are followed up ensure continuous improvement. Results are communicated to all staff

(6) Performance dialogue Score 1 Score 3 Score 5 Scoring grid: The right data or information for a

constructive discussion is often not present or conversations overly focus on data that is not meaningful. Clear agenda is not known and purpose is not stated explicitly

Review conversations are held with the appropriate data and information present. Objectives of meetings are clear to all participating and a clear agenda is present. Conversations do not, as a matter of course, drive to the root causes of the problems.

Regular review/performance conversations focus on problem solving and addressing root causes. Purpose, agenda and follow-up steps are clear to all. Meetings are an opportunity for constructive feedback and coaching.

(7) Consequence management Score 1 Score 3 Score 5 Scoring grid: Failure to achieve agreed objectives does

not carry any consequences Failure to achieve agreed results is tolerated for a period before action is taken.

A failure to achieve agreed targets drives retraining in identified areas of weakness or moving individuals to where their skills are appropriate

(8) Target balance Score 1 Score 3 Score 5 Scoring grid: Goals are exclusively financial or

operational Goals include non-financial targets, which form part of the performance appraisal of top management only (they are not reinforced throughout the rest of organization)

Goals are a balance of financial and non-financial targets. Senior managers believe the non-financial targets are often more inspiring and challenging than financials alone.

(9) Target interconnection Score 1 Score 3 Score 5

44

Scoring grid: Goals are based purely on accounting figures (with no clear connection to shareholder value)

Corporate goals are based on shareholder value but are not clearly communicated down to individuals

Corporate goals focus on shareholder value. They increase in specificity as they cascade through business units ultimately defining individual performance expectations.

(10) Target time horizon Score 1 Score 3 Score 5 Scoring grid: Top management's main focus is on short

term targets There are short and long-term goals for all levels of the organization. As they are set independently, they are not necessarily linked to each other

Long term goals are translated into specific short term targets so that short term targets become a "staircase" to reach long term goals

(11) Targets are stretching Score 1 Score 3 Score 5 Scoring grid: Goals are either too easy or impossible to

achieve; managers provide low estimates to ensure easy goals

In most areas, top management pushes for aggressive goals based on solid economic rationale. There are a few "sacred cows" that are not held to the same rigorous standard

Goals are genuinely demanding for all divisions. They are grounded in solid, solid economic rationale

(12) Performance clarity Score 1 Score 3 Score 5 Scoring grid: Performance measures are complex and not

clearly understood. Individual performance is not made public

Performance measures are well defined and communicated; performance is public in all levels but comparisons are discouraged

Performance measures are well defined, strongly communicated and reinforced at all reviews; performance and rankings are made public to induce competition

(13) Managing human capital Score 1 Score 3 Score 5 Scoring grid: Senior management do not communicate

that attracting, retaining and developing talent throughout the organization is a top priority

Senior management believe and communicate that having top talent throughout the organization is a key way to win

Senior managers are evaluated and held accountable on the strength of the talent pool they actively build

(14) Rewarding high-performance Score 1 Score 3 Score 5 Scoring grid: People within our firm are rewarded

equally irrespective of performance level Our company has an evaluation system for the awarding of performance related rewards

We strive to outperform the competitors by providing ambitious stretch targets with clear performance related accountability and rewards

45

(15) Removing poor performers Score 1 Score 3 Score 5

Scoring grid: Poor performers are rarely removed from their positions

Suspected poor performers stay in a position for a few years before action is taken

We move poor performers out of the company or to less critical roles as soon as a weakness is identified

(16) Promoting high performers Score 1 Score 3 Score 5

Scoring grid: People are promoted primarily upon the basis of tenure

People are promoted upon the basis of performance

We actively identify, develop and promote our top performers

(17) Attracting human capital Score 1 Score 3 Score 5

Scoring grid: Our competitors offer stronger reasons for talented people to join their companies

Our value proposition to those joining our company is comparable to those offered by others in the sector.

We provide a unique value proposition to encourage talented people join our company above our competitors

(18) Retaining human capital Score 1 Score 3 Score 5

Scoring grid: We do little to try to keep our top talent. We usually work hard to keep our top talent. We do whatever it takes to retain our top talent.

Source: Bloom and Van Reenen (2007)

46

TABLE A2: SAMPLING FRAME SOURCES AND SAMPLE RESPONSE RATES BY COUNTRY

(1) (2) (3) (4) (5) (6)

CODE Name Primary Source % responding #Contacted & eligible Years

AR Argentina BVD ORBIS 42.26 892 2010; 2013 AU Australia Dun & Bradstreet 33.24 1357 2009 BR Brazil BVD ORBIS 45.45 1727 2008;2013 CA Canada BVD ORBIS 34.4 1218 2008 CL Chile ENIA 70.55 550 2009 CN China BVD ORBIS 43.91 1273 2007;2010 CO Colombia Business Ministry 49.28 345 2013 ES Spain BVD ORBIS 29.8 708 2013 ET Ethiopia IGC Enterprise Map 60.65 216 2013 FR France BVD ORBIS 41.93 1047 2004;2006;2009;2014 GB Great Britain BVD ORBIS 29.2 2661 2004;2006;2009;2014 GE Germany BVD ORBIS 33.81 1121 2004;2006;2009;2014 GH Ghana IGC Enterprise Map 78.69 122 2013 GR Greece BVD ORBIS 53.96 782 2006;2010;2014 IN India CMIE FIRSTSOURCE 2005 43.42 1177 2006,2008;2010 IR Ireland BVD ORBIS 42.97 384 2008 IT Italy BVD ORBIS 48.31 888 2006;2010;2014 JP Japan BVD ORBIS 17.33 554 2006;2010 KE Kenya BVD ORBIS 66.91 272 2013 MM Myanmar PEDL Enterprise Map 65.04 226 2014 MX Mexico BVD ORBIS 39.81 967 2010;2013 MZ Mozambique IGC Enterprise Map 81.06 132 2013 NC Nicaragua Business Directories 80.2 101 2013 NG Nigeria BVD ORBIS 90.98 122 2013 NI Northern Ireland BVD ORBIS 62.98 181 2008 NZ New Zealand Dun & Bradstreet 44.19 344 2009 PO Poland BVD ORBIS 37.44 649 2006;2010 PT Portugal BVD ORBIS 53.72 363 2006;2010;2014 SG Singapore Trade & Industry Ministry 66.08 398 2012 SW Sweden BVD ORBIS 40.42 819 2006;2010 TR Turkey BVD ORBIS 81.92 177 2013 TZ Tanzania IGC Enterprise Map 29.14 3092 2013 US United States BVD ORBIS 38.92 388 2004;2006;2009;2014 VN Viet Nam BVD ORBIS 64.36 101 2014 ZM Zambia IGC Enterprise Map 42.26 892 2013 ALL 40.93 25,345

Notes: This tables contains some descriptive statistics of the countries in our database and the primary source of the sampling frame (more details are in Appendix A). Column (3) is the response rate (proportion of eligible and contacted firms who responded). “All” in the last row is the average across the entire sampling frame – i.e. weighted by number of firms contacted rather than simple unweighted average across countries (not including Singapore). Final column is main years included in survey.

47

TABLE A3: DISTRIBUTION OF WORKERS IN DIFFERENT FIRM SIZE CLASSES ACROSS COUNTRIES

% workers in firms with: France Germany Greece Ireland Italy Japan Poland Portugal Sweden GB US Under 50 employees 30.1% 17.8% 41.4% 28.3% 45.2% 23.9% 27.2% 51.4% 23.8% 36.2% 16.2% Between 50 and 5,000 employees 48.6% 52.8% 53.9% 71.7% 48.6% 58.6% 71.0% 47.8% 54.4% 49.3% 49.1% Over 5,000 employees 21.3% 29.4% 4.6% 0.0% 6.2% 17.5% 1.8% 0.7% 21.9% 14.5% 34.7%

Notes: This table displays estimates of the distribution of employment in manufacturing across firms in different size classes. The WMS sampling frame covers medium sized firms (between 50 and 5,000 workers) which usually covers half or more of total workers. The US (2006) and Japanese (2007) data are from published Census Bureau data and UK data (2010) is our analysis of unpublished Census data. For the other countries we use Eurostat data (which is based on Census) for the proportion of employment in firms with under 50 employees. For disclosure reasons, the proportion of employees in firms in over 5,000 employees is not reported in public use tables, however. Consequently, we used other data sources to estimate this fraction since we know the total manufacturing employment and we have access to the employment of the largest firms in every country from ORBIS company accounts data. Details are in Appendix B.

48

TABLE A4: RESPONSE RATES TO SURVEY

Dependent Variable: Responded = 1 (0 otherwise) (1) (2) (3) (4) (5) (6) Ln(Employment) 0.062*** 0.069*** 0.064*** 0.062*** 0.063*** 0.059*** (0.010) (0.010) (0.010) (0.010) (0.010) (0.011) Ln(Labor Productivity) 0.008 (0.010) Return on Capital Employed -0.013 (0.011) Multinational subsidiary? 0.009 (0.026) Ln(Firm age) 0.014 (0.013) Ln(Capital) 0.013 (0.009) # firms/observations 24,921 24,921 24,921 24,921 24,921 24,921

Notes: ***denotes significance at the 1% level, ** 5% level and *10% level. Marginal effects and standard errors from Probit ML. The dependent variable =1 if the firm responded to our survey and zero otherwise. An observation is the first a firm was contacted (so subsequent panel observations are not included, not are multiple plants in the same year which is why there are fewer observations than in the entire database). Years are 2006 to 2014 (see Bloom and Van Reenen, 2007, for analysis of the 2004 data). All regressions include three digit industry dummies, country dummies and year dummies (for survey response year and year that accounting data is taken from). Missing values on accounting variables are set to zero and a dummy for missing value included in the regression. Number of observations is 1,534 less than Table A1 because Singapore not included since we do not have details on covariates of non-responders from sampling frame.

49

TABLE A5: MANAGEMENT AND SUPPLY OF SKILLS (1) (2) (3) (4)

Dependent variable: Management % Employees with college Management Management

Method OLS OLS OLS IV Drive time to nearest -0.044*** -1.634*** University/B-School (0.016) (0.359) % Employees with 0.008*** 0.027*** college in the firm (0.001) (0.008) Observations 6,406 6,406 6,406 6,406

Notes: *** significance at the 1%, 5% (**) or 10% (*) level. Estimates by OLS with standard errors clustered by (313) sub-national regions. All columns include local population density, distance to coast, weather, firm age, skills, noise controls (interviewer dummies, reliability score, the manager’s seniority and tenure and the duration of the interview), a full set of three digit industry dummies, regional and year dummies, and a full set of country dummies. Management is a z-score of the average z-scores of the 18 management questions. See Feng (2013) for more details on distance measures. IV in column (4) is the drive time to nearest university

2 2.2 2.4 2.6 2.8 3 3.2 3.4

United StatesGermany

SwedenCanada

United KingdomFrance

ItalyAustralia

SingaporeMexicoPoland

PortugalChinaChile

GreeceSpainBrazil

ArgentinaIndia

Notes: Unweighted means of management scores by companies belonging to foreign multinationals vs. domestic firms (including domestic multinationals). Countries with 60 or more + foreign multinationals only (sample of 13,544 interviews).

Foreign multinationalsDomestic firms

Figure A1: Foreign multinationals vs. domestic firms

Management Score

910

1112

13Lo

g Sa

les

1 2 3 4 5Management

Figure A2: Firm Size is increasing in management

Notes: This plots the lowess predicted values of log(sales) against management (bandwidth=0.5). Sales is log(sales) in US$. N=13,854

-.4-.2

0.2

.4lo

wes

s: R

esid

uals

1 2 3 4 5Management

Figure A3: Firm TFP is increasing in management

Notes: This plots the lowess predicted valued of TFP against management (bandwidth=0.5). TFP calculated asresidual of regression of ln(sales) on ln(capital) and ln(labor) plus a full set of 3 digit industry, country and yeardummies controls. N = 10,900.

0 20 40 60 80 100

SwedenCanadaGreat BritainFranceGermanyAustraliaItalyPolandNew ZealandSpainMexicoJapanGreeceChileArgentinaPortugalColombiaIndiaBrazilChinaMozambiqueTanzaniaKenyaGhanaZambia

Figure A4: TFP Gap accounted for by management

Notes: TFP gaps from Penn World Tables; fraction accounted for by management uses the weighted averagemanagement scores and an assumed 10% impact of management on TFP. See Table 7 for details ofcalculations. Vertical line indicated the mean effect (30%).

% of TFP gap accounted for by managementTFP level (Relative to US)